{"version":"1","families":[{"slug":"claude","name":"Claude","provider":"Anthropic","status":"benchmarked","bestFor":"Polished apps that keep improving","verdict":"Claude produces the most polished finished apps we test and handles follow-up changes especially well. Its main risk is losing details in a long request. Opus 5.5, the newest, shared the top result of our latest phone-form test.","availability":"Claude is not in the Taskade model picker. It runs on your own Claude API key, in automations and in custom agents on Pro and up.","axes":{"Interface":"Strong","Task":"Strong","Memory":"Emerging","Adapt":"Leading"},"score":75,"variants":[{"anchor":"opus-5-5","name":"Claude Opus 5.5","asOf":"Sep 2026","verdict":"The newest Claude we test. On a family intake form it saved every answer and both photos, turned away an oversized photo before the upload, and was the only build of its test to open in the warm cream look the request asked for.","evidence":[{"date":"2026-09-22","claim":"Asked where the weekly recap should go and who would use the tracker before it built, then built three pages and a coaching automation that wrote a note into a logged match, at a far higher cost than the economical models."},{"date":"2026-09-30","claim":"Built the family intake form so it saved every answer and both photos, and it turned away a 12 MB photo before the upload started."},{"date":"2026-09-30","claim":"The only build of its test that opened in the light cream look the request asked for."},{"date":"2026-09-30","claim":"Added the new teacher question and put the change live. Its class assistant answered questions about the class but could not count the families."}]},{"anchor":"sonnet","name":"Claude Sonnet 5","asOf":"Oct 2026","verdict":"Our most carefully finished result in August. In September, one dashboard could not read its own sample data and a longer build stopped before the app was written. Its latest build, a family intake form, saved every answer and both photos.","evidence":[{"date":"2026-07-30","claim":"Opened its own app and checked the assistant's answers: the only model to verify its work this way."},{"date":"2026-08-03","claim":"Most carefully finished result, with eight minor items left."},{"date":"2026-08-03","claim":"Built a contacts database instead of the requested sign-up form after two stalls."},{"date":"2026-08-06","claim":"Four attempts stalled before the sign-up form was completed."},{"date":"2026-08-25","claim":"Stopped after asking its questions on the client sign-up form, and wrote the plan instead of the app."},{"date":"2026-09-19","claim":"Typed an 81-row scoring rubric and a 36-field sign-up table, then stopped at an approval prompt before writing the app."},{"date":"2026-09-19","claim":"Built a three-page match tracker whose dashboard reported zero matches over the eight matches it had entered itself."},{"date":"2026-09-30","claim":"Built the family intake form so it saved every answer and both photos, then added the new teacher question and put the change live."},{"date":"2026-10-08","claim":"Built the family intake form twice. Both builds saved every answer and both photos and took the new teacher question, and both showed sample families on the teacher page with no label."},{"date":"2026-10-08","claim":"Built the Dota 2 match tracker twice, each with a win-rate trend on the dashboard, and both saved a logged match. One ran wider than a phone screen. In the other the change added a second heroes page beside the one it had already built."}]},{"anchor":"sonnet-4-5","name":"Claude Sonnet 4.5","asOf":"Oct 2026","verdict":"An older Claude, tested in October. Its match tracker rejected every logged match and read every sample match as a loss, while its family intake form saved every answer and both photos.","evidence":[{"date":"2026-10-08","claim":"Built the family intake form so it saved every answer and both photos, in a plain look with little of the warm pastel the request asked for. Its open teacher page listed parent phone numbers."},{"date":"2026-10-08","claim":"Built a Dota 2 match tracker that rejected every logged match and showed every sample match as a loss, so the dashboard read a 0% win rate."}]},{"anchor":"opus","name":"Claude Opus 5","asOf":"Sep 2026","verdict":"The most complete client sign-up build of September: five pages, every question in the customer's wording, and a scoring automation that ran and saved its verdict.","evidence":[{"date":"2026-07-30","claim":"Set the design high-water mark with a polished, consistent interface."},{"date":"2026-08-01","claim":"Created the fullest workspace, with an assistant that understood its contents."},{"date":"2026-08-25","claim":"The most complete client sign-up build of the test: a five-page form, every answer saved, and the scoring automation graded the submission on its own."},{"date":"2026-08-25","claim":"Best-executed match tracker of the test, without a single failed build action, at by far the highest cost."},{"date":"2026-09-19","claim":"Finished the client sign-up form with all 32 questions in the customer's wording, 31 word for word, five pages, and a scoring automation that saved its result."},{"date":"2026-09-19","claim":"Decided two open points in the brief and named both decisions in its closing summary."},{"date":"2026-09-30","claim":"Built the family intake form so it saved every answer and both photos, and its email automation ran three times. It cost far more than Claude Opus 5.5 on the same request for a lower result."}]},{"anchor":"haiku","name":"Claude Haiku 4.5","asOf":"Oct 2026","verdict":"Caught and repaired a problem that would have left the app blank.","evidence":[{"date":"2026-07-31","claim":"Found the problem while building, repaired it, and then finished: the only model in the test to recover on its own."},{"date":"2026-08-25","claim":"Fastest build of the test, but it left out the weekly recap automation the brief asked for."},{"date":"2026-10-08","claim":"Built the family intake form in under 4 minutes and took the new teacher question, but the family photo replaced the child photo on every submission."},{"date":"2026-10-08","claim":"Built a Dota 2 match tracker that saved logged matches and added a heroes page when asked, but its top-heroes chart showed empty axes."}]}],"faq":[{"q":"What is Claude best at in Taskade Genesis?","a":"Claude is strongest when finish and follow-up changes matter. Sonnet produced our most carefully finished result, while Opus set the visual quality high-water mark. With long requests, review the finished app to make sure every detail carried through."},{"q":"Which Claude model is best for app building?","a":"Claude Opus 5.5 shared the top result of our latest phone-form test and opened in the look the request asked for. Claude Opus 5 produced the more complete September sign-up app, with five pages and a scoring automation that ran. Claude Sonnet 5 remains strong at careful finish and follow-up changes, but review its dashboards against the saved data."},{"q":"Why did Claude fail a test?","a":"In one August test it built a contacts database instead of the client sign-up form, after two three-minute stalls lost it the thread. In another, four stalled runs in a row meant the client sign-up form never got built at all. We publish the tests that go badly alongside the ones that go well, and we open every app and read it back against the brief. That is the point of the benchmark."},{"q":"Can I use Claude in Taskade?","a":"Yes, with your own Claude API key: use it in the Anthropic Claude automation step, or connect it to your custom agents on Pro and up. Claude is not in the Taskade model picker."},{"q":"Where can I compare Claude with other models, or try it myself?","a":"Side-by-side model pages live at /compare. Claude is not in the model picker, so /create starts a build with Auto or another available model. This page stays on what Claude produced in TSK-1 tests."}]},{"slug":"gpt","name":"GPT","provider":"OpenAI","status":"benchmarked","bestFor":"Detailed requests and fast delivery","verdict":"GPT is strongest at following detailed requests and finishing quickly. GPT-6.1 Sol, the newest, shared the top result of our latest phone-form test. Luna preserves exact wording and builds complete apps efficiently. Terra is the fastest in nearly every test it enters. GPT-5.6 Sol builds wide apps but can add sign-in screens nobody asked for.","axes":{"Interface":"Strong","Task":"Leading","Memory":"Strong","Adapt":"Leading"},"score":88,"variants":[{"anchor":"gpt-6-1-sol","name":"GPT-6.1 Sol","asOf":"Sep 2026","verdict":"The newest GPT we test. All three of its family intake forms saved every answer and both photos, parents could use them without an account as the request asked, and every follow-up change landed.","evidence":[{"date":"2026-09-30","claim":"Built the family intake form three times. All three saved every answer and both photos, and parents could fill them in without an account, as the request asked."},{"date":"2026-09-30","claim":"In two of the three builds the teacher page listed every family with their photos. In the third it read the saved answers wrong and showed \"No photos saved\" for a complete entry."},{"date":"2026-09-30","claim":"Added the new teacher question in all three apps, saved it with each submission, and put the change live."},{"date":"2026-09-30","claim":"Every closing summary said what it had not tested and pointed out that the teacher page was open to anyone with the link."}]},{"anchor":"gpt-6-luna","name":"GPT-6 Luna","asOf":"Oct 2026","verdict":"The newest Luna, tested on the Dota 2 match tracker. In October it built the tracker with every requested detail, a coach and a weekly recap in about 10 minutes, then added a heroes page when asked.","evidence":[{"date":"2026-09-22","claim":"Built a match tracker that saved a logged match and took any hero name, but its dashboard opened empty because the app hid every sample match. The build took about 18 minutes."},{"date":"2026-10-10","claim":"Built the Dota 2 match tracker in about 10 minutes with every detail the request named (hero, result, kills, deaths, assists, duration, notes, win rate and streaks), a coach agent and a weekly recap automation."},{"date":"2026-10-10","claim":"Added the heroes page and its menu link in about 8 minutes and put the change live. A logged match saved and showed in the app."}]},{"anchor":"luna","name":"GPT-5.6 Luna","asOf":"Oct 2026","verdict":"The most faithful model we have tested. It reproduced all 32 questions of a real customer's client sign-up form word for word again, built a scoring automation that ran, and passed every September follow-up-edit test.","evidence":[{"date":"2026-08-03","claim":"Reproduced all 32 questions of a real customer's client sign-up form word for word. The only model to get all four of its builds word for word across more than one test."},{"date":"2026-08-07","claim":"The only model to build a working scoring grid: 33 rows by 5 ratings, with every score saved straight back into the workspace."},{"date":"2026-08-04","claim":"The best client sign-up form Luna has built: nothing stalled, and a human reviewer called it beautiful."},{"date":"2026-08-08","claim":"Test winner on value. The only model to explain its own design choices, and both builds carried all 32 questions word for word."},{"date":"2026-08-19","claim":"The baseline for automatic model picking: an app built in about 11 minutes and edited in about 6."},{"date":"2026-08-25","claim":"Its match tracker crashed on the first real entry and its sign-up form rejected the first submission; later tests finished cleanly."},{"date":"2026-08-27","claim":"The value pick on every app in a five-model test of identical prompts."},{"date":"2026-08-28","claim":"Unchanged result after a large platform update, with fewer failed build actions."},{"date":"2026-08-29","claim":"Kept a build checklist inside both apps, so the finished work could be checked against the request."},{"date":"2026-09-19","claim":"Built the client sign-up form with four pages and a scoring automation that ran on its own."},{"date":"2026-09-19","claim":"Built the fleet-inspection app with both weekly automations live in 7 minutes 18 seconds."},{"date":"2026-09-19","claim":"Built the AI governance office in 6 minutes 38 seconds with all 24 named items in place."},{"date":"2026-09-19","claim":"Built a complete match tracker with a live dashboard and a weekly recap automation."},{"date":"2026-09-19","claim":"Passed all five follow-up-edit tests on its habit tracker."},{"date":"2026-09-19","claim":"Its recipe box rendered as a warm printed cookbook in light and dark, with every page loading cleanly."},{"date":"2026-09-22","claim":"Built a two-page match tracker in 8 minutes 19 seconds, the fastest and most economical complete build of its test, and a match saved from the app came back with all eight details in place."},{"date":"2026-09-30","claim":"Built the family intake form so it saved every answer and both photos, and added the new teacher question cleanly: the most economical complete build of its test."},{"date":"2026-09-30","claim":"Its public class assistant told an anonymous visitor a child's name from a saved form."},{"date":"2026-10-08","claim":"Built the family intake form twice. One build saved every answer and both photos in under 5 minutes. In the other, every parent submission failed on the photo permission question, although its closing summary said the form had worked."},{"date":"2026-10-08","claim":"Built the Dota 2 match tracker twice in about 5 and a half minutes each. Both saved a logged match with every detail and added a ranked heroes page when asked."}]},{"anchor":"terra","name":"GPT-5.6 Terra","asOf":"Oct 2026","verdict":"The fastest model in nearly every test it enters, and the first to pass every check in a single run: it built the app, took a form submission, put every answer in the right place, ran the automation, and answered questions about the data. It also turned its weakest result around, from three tests where none of the brief came through word for word to all 32 questions word for word.","evidence":[{"date":"2026-08-07","claim":"The first model to pass all five checks in one run: built the app, took a submission, saved every answer in the right field, ran the automation, and answered questions about the data."},{"date":"2026-08-07","claim":"Fastest of the test: 8 minutes 26 seconds."},{"date":"2026-08-08","claim":"Both builds carried all 32 questions word for word, after three tests where none of the four came through. A change of setting turned it around."},{"date":"2026-08-08","claim":"Terra's premium setting was the worst value we have measured on the client sign-up form: far more expensive than Luna, and it delivered less, a single page rewritten 19 times."},{"date":"2026-08-25","claim":"Best all-round match tracker and best client sign-up build of the test: form to saved score worked on the first try."},{"date":"2026-08-27","claim":"The quality pick in a five-model test of identical prompts."},{"date":"2026-09-19","claim":"Fastest complete client sign-up build of the test at 5 minutes 6 seconds; it left the scoring automation out."},{"date":"2026-09-19","claim":"Finished a two-page sales pipeline with a daily call brief in 10 minutes 47 seconds."},{"date":"2026-09-19","claim":"Finished the governance office in under 6 minutes and the fleet inspection in under 5."},{"date":"2026-09-19","claim":"Passed all five follow-up-edit tests; its recipe box was a single page."},{"date":"2026-09-22","claim":"Built a clean match tracker in about 11 minutes, but its match form offered only four heroes from its own sample list, so a player could not log the hero they played."},{"date":"2026-10-08","claim":"Built the family intake form twice, about 4 minutes each, and both saved every answer and both photos. One opened in a dark look where the request asked for cream paper. The other asked first who could open the teacher page, then put it behind a teacher sign-in."},{"date":"2026-10-08","claim":"Built the Dota 2 match tracker twice in under 4 and a half minutes, and both added a ranked heroes page when asked. One build labeled its starter matches, and the other ran wider than a phone screen on every page."}]},{"anchor":"sol","name":"GPT-5.6 Sol","asOf":"Oct 2026","verdict":"The login-wall lesson, and a turnaround. Through September Sol built wide apps with complete data but kept adding sign-in screens nobody asked for. In October its family intake form needed no sign-in, both of its apps labeled their sample data, and both follow-up changes landed.","evidence":[{"date":"2026-07-30","claim":"Added a sign-in screen the brief never asked for, and did not say so."},{"date":"2026-08-01","claim":"Added the same unasked-for sign-in screen again. Repeatable enough that it shaped how we score a build that answers a different brief than the one given."},{"date":"2026-08-25","claim":"Added a sign-in gate to the match tracker again, the third time in three tests."},{"date":"2026-09-19","claim":"Put an unasked sign-in screen in front of two of its three finished client sign-up builds, over complete data and three scoring runs."},{"date":"2026-09-19","claim":"Built a 34-field fleet-inspection form and left the Thursday reminder switched off."},{"date":"2026-09-19","claim":"Built the widest governance office of the test, with seven pages and every named item, behind a sign-in screen."},{"date":"2026-09-19","claim":"Built a two-page match tracker with no sign-in screen."},{"date":"2026-09-19","claim":"Passed all five follow-up-edit tests, but the habit tracker sat behind a sign-in screen it added itself."},{"date":"2026-09-22","claim":"Put a sign-in screen nobody asked for in front of a personal match tracker, and its summary did not mention that an account had to be created first."},{"date":"2026-10-08","claim":"Built the family intake form so it saved every answer and both photos, labeled its sample family, and said on the teacher page that it needs no sign-in. The new teacher question landed."},{"date":"2026-10-08","claim":"Built the Dota 2 match tracker with its sample matches labeled as a demo, saved a logged match with every detail, and put the new heroes page in both the desktop and the phone menu."}]}],"faq":[{"q":"Which GPT variant builds the best apps?","a":"GPT-6.1 Sol, the newest, saved every answer and photo in all three of its phone-form builds and landed every follow-up change. GPT-5.6 Luna sticks to the brief best and builds complete apps efficiently. GPT-5.6 Terra is the fastest and was the first to pass every check in one run. GPT-5.6 Sol builds wider apps, but it can add sign-in screens nobody asked for."},{"q":"Is a faster AI model worse for app building?","a":"Not necessarily. GPT-5.6 Terra is the fastest model we test and the first to pass every check in one run. But speed only counts if the app matches the brief. For three tests Terra reproduced none of the customer's questions word for word; a change of setting brought it back to all 32."},{"q":"Can I use GPT models in Taskade?","a":"Yes. GPT-6.1 Sol, GPT-6 Luna, and GPT-5.6 Luna are in the Taskade model picker on every paid plan. On Max and Enterprise, you can also pick GPT-5.6 Sol and GPT-5.6 Terra directly."},{"q":"Where can I compare GPT with other models, or try it myself?","a":"Side-by-side model pages live at /compare. To try the same kind of app request yourself, start from /create. This page stays on what GPT produced in TSK-1 tests."}]},{"slug":"deepseek","name":"DeepSeek","provider":"DeepSeek","status":"benchmarked","bestFor":"Polished apps with rich workspace data","verdict":"DeepSeek combines rich apps with efficient builds. V4.1 Flash built the richest sales CRM and cookbook in their tests. Pro builds deep data but can leave the app locked or unfinished. In October both built a family intake form that saved every answer and both photos.","axes":{"Interface":"Leading","Task":"Strong","Memory":"Strong","Adapt":"Emerging"},"score":75,"variants":[{"anchor":"flash","name":"DeepSeek V4 Flash","asOf":"Aug 2026","verdict":"The cheapest model won the design test. It was the only one that looked right in both light and dark, ran without a single error, and laid out cleanly on a phone, for a fraction of what the others cost. On a real customer brief it was cheapest, fastest and cleanest at once. It has since left the Taskade model picker.","evidence":[{"date":"2026-08-01","claim":"Won on design: the only model that looked right in both light and dark, ran without a single error, and laid out cleanly on a 390px phone screen, at a fraction of the field's cost."},{"date":"2026-08-02","claim":"Won on design again, and this time the finished workspace saved real data with light and dark themes both working throughout."},{"date":"2026-08-03","claim":"Cheapest, fastest and cleanest run on the 32-question client sign-up form: 12 minutes 36 seconds."},{"date":"2026-08-02","claim":"Same test, same prompt: better value than Luna, at half the cost of the build."},{"date":"2026-08-25","claim":"Finished the tracker and made both follow-up changes cleanly, but its long-form build ran out of time."}]},{"anchor":"flash-4-1","name":"DeepSeek V4.1 Flash","asOf":"Oct 2026","verdict":"The richest app per shape among the economical models. It built a six-page sales CRM whose scoring automation ran, a four-page match tracker, and a four-page cookbook with both automations live. On a family intake form it saved every answer and both photos.","evidence":[{"date":"2026-09-19","claim":"Built the richest sales CRM of the test, with six pages and a scoring automation that scored all 7 sample leads."},{"date":"2026-09-19","claim":"Built a four-page match tracker and said plainly that its weekly recap was still switched off."},{"date":"2026-09-19","claim":"Typed all 32 client sign-up questions and built a 48-field submissions table, then ran out of time before finishing."},{"date":"2026-09-19","claim":"Built a three-page fleet inspection app for 42 vans where the brief named 40."},{"date":"2026-09-19","claim":"Its recipe box came out as a printed cookbook with 13 recipes, four pages, and both automations live."},{"date":"2026-09-22","claim":"Built a three-page match tracker whose coaching automation ran on the first match logged and wrote a note back into that match about six minutes later."},{"date":"2026-09-30","claim":"Built a two-page family intake form that saved every answer and both photos, and its alert automation ran twice. The build ran past the 15-minute test limit."},{"date":"2026-09-30","claim":"Added the new teacher question and put the change live."},{"date":"2026-10-08","claim":"Built the family intake form in 7 minutes 25 seconds so it saved every answer and both photos, and its new-family automation ran. Three sample families showed on the teacher page as if they were real."},{"date":"2026-10-08","claim":"Added the new teacher question as an optional field and kept it off the printed facesheet, and said so."},{"date":"2026-10-08","claim":"Built a five-page Dota 2 match tracker with labeled sample matches, but every hero showed as unknown and no match could be saved, while its summary said the save had been checked."},{"date":"2026-10-08","claim":"Rebuilt the 32-question client sign-up form with all 32 questions word for word and a scoring rubric, but the form saved each answer one question off, so every submission was rejected."}]},{"anchor":"pro","name":"DeepSeek V4 Pro","asOf":"Oct 2026","verdict":"The workspace champion with a delivery problem. Pro still builds deep data. In September it put two apps behind a sign-in screen and left a third without a page. In October it built a family intake form that saved every answer and both photos, with no sign-in.","evidence":[{"date":"2026-08-06","claim":"The richest workspace of the test: 8 automations, a record 60 fields, and all four builds word for word."},{"date":"2026-08-05","claim":"Won the tracker test as the most efficient app that opened. Two stronger-looking results could not be used, so they did not receive a score."},{"date":"2026-08-02","claim":"Went from none of its four builds matching the brief to all four word for word on a shorter prompt. The biggest jump we have measured in a single test."},{"date":"2026-08-25","claim":"The cheapest complete match tracker of the test."},{"date":"2026-09-19","claim":"Built the deepest lead table of the sales CRM test, with 12 fields per lead, then put its one-page board behind a sign-in screen."},{"date":"2026-09-19","claim":"Finished the client sign-up form with all 32 questions word for word and left both automations switched off."},{"date":"2026-09-19","claim":"Spent the fleet-inspection test building data and automations and wrote no page."},{"date":"2026-09-19","claim":"Built a complete match-tracker data layer, then put the app behind a sign-in screen."},{"date":"2026-09-22","claim":"Built one of the two best-looking match trackers of its test, with a win-rate trend chart in light and dark, in 36 minutes, but its hero picker listed only 17 heroes."},{"date":"2026-10-08","claim":"Built the family intake form in 7 minutes 24 seconds so it saved every answer and both photos, and the new teacher question landed. Two sample families showed as real, and its summary did not mention them."},{"date":"2026-10-08","claim":"Built a Dota 2 match tracker with a win-rate trend that saved a logged match and added the heroes page when asked. About 15 sample matches showed as real."}]},{"anchor":"pro-max-thinking","name":"DeepSeek V4 Pro, max thinking","asOf":"Oct 2026","verdict":"DeepSeek V4 Pro with its deepest thinking setting, which the Taskade model picker no longer offers. The 15-minute test limit ended its build before a closing summary, and the follow-up turn finished an intake form that saved every answer and both photos with no fake sample families.","evidence":[{"date":"2026-10-08","claim":"The 15-minute test limit ended the family intake build before a closing summary. The follow-up turn finished the app and added the new teacher question."},{"date":"2026-10-08","claim":"The finished form saved every answer and both photos, and the teacher page started empty instead of showing made-up families."},{"date":"2026-10-08","claim":"Built the best-looking Dota 2 match tracker of its set, and a logged match saved. The 15-minute limit ended the build, and the follow-up turn finished the interface and the heroes page."}]},{"anchor":"pro-high-thinking","name":"DeepSeek V4 Pro, high thinking","asOf":"Oct 2026","verdict":"DeepSeek V4 Pro with a high thinking setting, which the Taskade model picker no longer offers. Its family intake form could never keep a photo, so no parent could submit it. Its match tracker labeled its demo matches and kept them out of the stats.","evidence":[{"date":"2026-10-08","claim":"The 15-minute test limit ended the family intake build before a closing summary, and the form dropped every photo a parent picked, so the two-photo rule blocked every submission."},{"date":"2026-10-08","claim":"Built a Dota 2 match tracker that saved a logged match, labeled its demo matches and kept them out of the stats, and added the heroes page when asked."}]},{"anchor":"pro-0813","name":"DeepSeek V4 Pro 0813","asOf":"Oct 2026","verdict":"A dated DeepSeek V4 Pro release that is not in the Taskade model picker. With thinking off and with max thinking, both of its family intake forms saved every answer and both photos, and the new question landed each time.","evidence":[{"date":"2026-10-08","claim":"With thinking off, it built the family intake form in 6 minutes 14 seconds so it saved every answer and both photos, and it removed its own test row when it finished."},{"date":"2026-10-08","claim":"With max thinking, the 15-minute test limit ended the build, and the follow-up turn finished a form that saved every answer and both photos, with every sample family labeled."},{"date":"2026-10-08","claim":"Built the Dota 2 match tracker with both settings, and both saved a logged match. With max thinking it labeled its sample matches. With thinking off it had already built a ranked heroes tab, so the follow-up change made no edit and said so."}]},{"anchor":"v3-2","name":"DeepSeek V3.2","asOf":"Oct 2026","verdict":"An older DeepSeek that is not in the Taskade model picker. With thinking off and on, neither of its family intake forms could save a family, though one uploaded both photos. V4.1 Flash, the DeepSeek Flash model tested here, saved every answer and photo on the same form.","evidence":[{"date":"2026-10-08","claim":"With thinking off, the form uploaded both photos, but the save was rejected, and the teacher page showed its sample families with blank names."},{"date":"2026-10-08","claim":"With thinking on, the form sent each submission to an address that does not exist, and the teacher page showed families typed into the page instead of saved ones."}]}],"faq":[{"q":"What is DeepSeek best at in Taskade Genesis?","a":"DeepSeek is strongest on workspace depth. V4 Flash won an August design test. V4.1 Flash built a six-page sales CRM with a scoring automation that ran. Pro still builds deep data, but its September apps were less reliable to reach."},{"q":"Which DeepSeek model is best for app building?","a":"In these tests, V4.1 Flash, the DeepSeek Flash model tested here, gave the richest app at a low cost: it built a six-page sales CRM and saved every answer and photo on a family intake form. Choose Pro when data depth matters most, then check that every page and automation is available. V4 Flash, which won an August design test, has left the picker."},{"q":"Are DeepSeek models open source?","a":"DeepSeek publishes open-weight models, which means the underlying weights are available rather than kept entirely private. You can pick DeepSeek models in the Taskade model picker."},{"q":"Where can I compare DeepSeek with other models, or try it myself?","a":"Side-by-side model pages live at /compare. To try the same kind of app request yourself, start from /create. This page stays on what DeepSeek produced in TSK-1 tests."}]},{"slug":"gemini","name":"Gemini","provider":"Google","status":"benchmarked","bestFor":"A fast-improving Google model","verdict":"Gemini improved between tests. An early result never opened; the next built, published, and worked. We use every finished app so progress is measured by what people can actually run.","availability":"Gemini is not in the Taskade model picker right now. It runs in your automations on your own Gemini API key.","axes":{"Interface":"Limited","Task":"Emerging","Memory":"Emerging","Adapt":"Emerging"},"score":44,"variants":[{"anchor":"flash-3-6","name":"Gemini 3.6 Flash","asOf":"Aug 2026","verdict":"An early Gemini result never opened: nothing to see and nothing to use. The next one built and published. That is real progress, confirmed by opening the app rather than taking the model's word for it.","evidence":[{"date":"2026-08-01","claim":"The finished build never opened: three missing pieces, no light-and-dark styling, and links pointing at a machine nobody else can reach. No app shipped."},{"date":"2026-08-03","claim":"Built and published this time, a clear step up. It still used too many resources to qualify as the best value in the test."}]}],"faq":[{"q":"What is Gemini best at in Taskade Genesis?","a":"Gemini 3.6 Flash showed clear improvement between tests: an early build never opened, and the next one built and published. We measure progress by opening every app, not by taking the model at its word."},{"q":"Does Gemini work in Taskade?","a":"Not in the model picker right now. Gemini was part of the benchmark, and its story is a simple one: an early build never opened at all, and the next one built and published. We open every app, which is the only way to tell \"I built it\" from \"it runs.\""},{"q":"Can I use Google models in Taskade?","a":"Not from the model picker right now. With your own Gemini API key, the Google Gemini automation step runs Gemini in your automations."},{"q":"Where can I compare Gemini with other models, or try a build myself?","a":"Side-by-side model pages live at /compare. Gemini is not in the picker right now, so /create starts a build with Auto or another available model. This page stays on what Gemini produced in TSK-1 tests."}]},{"slug":"glm","name":"GLM","provider":"Zhipu","status":"benchmarked","bestFor":"Good judgment when details are unclear","verdict":"The model that said no. It considered adding a sign-in screen, decided the brief had not asked for one, and offered it as a suggestion instead. The only model to push back rather than quietly add something. Its apps have saved real data since an August test.","axes":{"Interface":"Emerging","Task":"Strong","Memory":"Emerging","Adapt":"Strong"},"score":63,"variants":[{"anchor":"glm-5-2","name":"GLM-5.2","asOf":"Sep 2026","verdict":"The model that said no in July, and a steady finisher in September. It built a live win-rate dashboard and a three-page sign-up app with 31 of 32 questions word for word.","evidence":[{"date":"2026-07-30","claim":"Thought about adding a sign-in screen, decided the brief had not asked for one, and offered \"Add Login\" as a suggestion instead. The only model to push back rather than quietly add it."},{"date":"2026-08-01","claim":"Its app saved real data for the first time. A small styling issue is the one thing still open."},{"date":"2026-08-25","claim":"Complete and working, but its sample matches carried no result, so the dashboard showed a 0% win rate until a real match was logged."},{"date":"2026-09-19","claim":"Built a two-page match tracker whose dashboard showed a 58% win rate over 12 sample matches: 7 wins and 5 losses."},{"date":"2026-09-19","claim":"Built a three-page client sign-up app with a 37-field submissions table and 31 of 32 questions word for word."}]},{"anchor":"glm-5-3","name":"GLM-5.3","asOf":"Aug 2026","verdict":"Built a richer app than most, then ran out of time before finishing it. On the long form it thought at length and never started. Both follow-up edits were clean. GLM-5.3 is now in the Taskade model picker, and this August test is its latest result.","evidence":[{"date":"2026-08-29","claim":"Four pages built, but the build did not finish inside the time limit."},{"date":"2026-08-29","claim":"On the 32-question client sign-up form it produced no app."},{"date":"2026-08-29","claim":"Both follow-up edits landed with no failed build actions."}]},{"anchor":"glm-5","name":"GLM-5","asOf":"Oct 2026","verdict":"Not in the Taskade model picker. No link led to the gallery page of its family intake form, but the form saved every answer and both photos, and the gallery showed the new teacher question.","evidence":[{"date":"2026-10-08","claim":"Built the family intake form in 12 minutes 39 seconds so it saved every answer and both photos, and its gallery page showed each family with both photos."},{"date":"2026-10-08","claim":"Added the new teacher question and showed it on the gallery page, but no link in the app led to that page."}]},{"anchor":"glm-4-7","name":"GLM-4.7","asOf":"Oct 2026","verdict":"An older GLM that is not in the Taskade model picker. Its teacher page crashed on load, and its family intake form saved every answer and both photos.","evidence":[{"date":"2026-10-08","claim":"Built the family intake form so it saved every answer and both photos, but the teacher page crashed on load, so the new teacher question never showed there."}]}],"faq":[{"q":"What is GLM best at in Taskade Genesis?","a":"GLM is strongest when a brief is silent on extras. GLM-5.2 was the only model to decline an unasked sign-in screen and offer it as a suggestion instead. GLM-5.3 builds richer apps on shorter briefs, but long forms can run out of time before finishing."},{"q":"GLM-5.2 or GLM-5.3 for app building?","a":"Both are in the Taskade model picker on Max and Enterprise. GLM-5.2 is strongest on judgment and clean follow-up edits. GLM-5.3 built a richer first app on a shorter brief, but its long-form build ran out of time in its August test."},{"q":"Is GLM open source?","a":"GLM-5.2 is an open-weight model from Zhipu, and you can pick it in the Taskade model picker."},{"q":"How does GLM compare to Claude and GPT in TSK-1?","a":"GLM-5.2's standout result is about judgment rather than raw power: it was the only model to decline a sign-in screen the brief never asked for, and to offer it as an option instead. When a brief is silent, that is the right call, and it is exactly what we watch for."},{"q":"Where can I compare GLM with other models, or try it myself?","a":"Side-by-side model pages live at /compare. To try the same kind of app request yourself, start from /create. This page stays on what GLM produced in TSK-1 tests."}]},{"slug":"kimi","name":"Kimi","provider":"Moonshot","status":"benchmarked","bestFor":"Rich apps that need patience","verdict":"Kimi builds more than almost any model and finishes later than all of them. Its September apps were rich, but three tests reached the 40-minute limit before a closing summary.","axes":{"Interface":"Strong","Task":"Strong","Memory":"Strong","Adapt":"Strong"},"score":75,"variants":[{"anchor":"k3","name":"Kimi K3","asOf":"Sep 2026","verdict":"Builds more than almost any model and finishes later than all of them. Its sales CRM scored all eight sample leads and its match tracker was the richest of its test.","evidence":[{"date":"2026-07-31","claim":"The most reliable model of the test: only 7.2% of its build actions went wrong, and it needed the fewest actions. Its closing summary accurately described the app it had built."},{"date":"2026-08-25","claim":"A good-looking, complete match tracker with a heroes page; we could not confirm a saved match from outside the app."},{"date":"2026-09-19","claim":"Built a complete client sign-up app with 31 of 32 questions word for word, then reached the 40-minute limit before a closing summary."},{"date":"2026-09-19","claim":"Its sales CRM scoring automation ran eight times and scored all 8 sample leads."},{"date":"2026-09-19","claim":"Built the richest match tracker of the test, with four pages, heroes, and recaps."},{"date":"2026-09-30","claim":"On a real customer's family intake form it wrote no app page within the 15-minute test limit."}]},{"anchor":"k2-6","name":"Kimi K2.6","asOf":"Aug 2026","verdict":"Tested once in August on the Dota 2 match tracker. It built a clean dashboard with a win and loss timeline, KDA trends and top heroes, plus a coach and a weekly recap.","evidence":[{"date":"2026-08-25","claim":"Built a clean match dashboard with a win and loss timeline, KDA trend lines and top heroes, a 25-hero reference, a coach and a weekly recap, in about 16 minutes. We could not confirm a saved match from outside the app."}]},{"anchor":"k2-7-code","name":"Kimi K2.7 Code","asOf":"Aug 2026","verdict":"Tested once in August on the Dota 2 match tracker. It finished a tracker with a match form and a dashboard, and closed with a clear summary of what it built.","evidence":[{"date":"2026-08-25","claim":"Built a match tracker with a log form and a dashboard in about 30 minutes, but its three sample matches showed as empty cards, and it sent an email on every logged match in place of the weekly recap the request asked for."}]}],"faq":[{"q":"What is Kimi K3 best at in Taskade Genesis?","a":"Kimi K3 is best for rich apps when you can allow a longer build. Its September sales CRM scored all eight sample leads, and its match tracker was the richest of its test."},{"q":"Can I use Kimi in Taskade?","a":"Yes. Kimi K3, Kimi K2.6 and Kimi K2.7 Code are in the Taskade model picker on Max and Enterprise. Kimi K3 has the most results on this page. Kimi K2.6 and Kimi K2.7 Code each took one August test."},{"q":"Where can I compare Kimi with other models, or try it myself?","a":"Side-by-side model pages live at /compare. To try the same kind of app request yourself, start from /create. This page stays on what Kimi produced in TSK-1 tests."}]},{"slug":"minimax","name":"MiniMax","provider":"MiniMax","status":"benchmarked","bestFor":"Not scored in this test","verdict":"MiniMax M3 was not scored because too many build actions failed and the app was not ready to use. We open every result before ranking it. A newer version can earn a fresh test.","axes":{"Interface":"Not scored","Task":"Not scored","Memory":"Not scored","Adapt":"Not scored"},"variants":[{"anchor":"m3","name":"MiniMax M3","asOf":"Aug 2026","verdict":"Not scored. Nearly half of its build actions failed, yet it reported the app as ready. We open and use every result before ranking it. A newer version can earn a fresh test.","evidence":[{"date":"2026-07-31","claim":"Not scored: 47.2% of its build actions failed, and it reported the app as ready regardless of what its own record showed."},{"date":"2026-08-25","claim":"Its August build never opened, and the run ended without a closing message."}]}],"faq":[{"q":"Why doesn't MiniMax have a score?","a":"MiniMax M3 was not scored because nearly half of its build actions failed and the app was not ready to use. Reliability is required before a model enters the ranking. MiniMax remains available in Taskade, and a newer version can earn a fresh test."},{"q":"Can I use MiniMax in Taskade?","a":"Yes. MiniMax M3 is in the Taskade model picker. Its TSK-1 score stays open until a newer version completes a hands-on app test that meets our reliability bar."},{"q":"Where can I compare MiniMax with other models, or try it myself?","a":"Side-by-side model pages live at /compare. To try the same kind of app request yourself, start from /create. This page stays on what MiniMax produced in TSK-1 tests."}]},{"slug":"qwen","name":"Qwen","provider":"Alibaba","status":"benchmarked","bestFor":"Value on long, detailed forms","verdict":"Qwen is the value newcomer. In its first hands-on test, Qwen 3.7 Plus finished the 32-question client sign-up form with every answer saved, at a fraction of the usual cost, and Qwen 3.7 Max built a clean match tracker. Follow-up changes are where it still stalls: two asked-for edits were started and not finished.","axes":{"Interface":"Emerging","Task":"Strong","Memory":"Emerging","Adapt":"Limited"},"score":50,"variants":[{"anchor":"qwen-3-7-plus","name":"Qwen 3.7 Plus","asOf":"Aug 2026","verdict":"Finished the full 32-question client sign-up form in its first test, with every answer saved, for a fraction of what most models spend. Its apps are plainer than the leaders', it kept less around the app, and both follow-up edits stopped before they were done.","evidence":[{"date":"2026-08-25","claim":"Took all 32 questions of the client sign-up form, saved every answer, and did it at a fraction of the usual cost."},{"date":"2026-08-25","claim":"Built a working match dashboard quickly, but with the thinnest workspace behind it of any model in the test."},{"date":"2026-08-25","claim":"Started both follow-up edits and finished neither: a new page was written but never linked into the app."},{"date":"2026-08-25","claim":"On a harder field inspection brief it left the inspector with nothing to do on day one: the checklist could not start until a weekly automation had run."}]},{"anchor":"qwen-3-7-max","name":"Qwen 3.7 Max","asOf":"Aug 2026","verdict":"A clean, complete match tracker on its first run. One rough edge: with no hero picked, the save button did nothing and said nothing.","evidence":[{"date":"2026-08-25","claim":"Complete match tracker with a clean dashboard on the first run; it saved a real match once a hero was picked."},{"date":"2026-08-25","claim":"With the hero left blank, the save button did nothing and said nothing: a first-use trap for a real user."}]}],"faq":[{"q":"What is Qwen best at in Taskade Genesis?","a":"Value on long, detailed forms. In its first hands-on test Qwen 3.7 Plus took a 32-question client sign-up form, saved every answer, and spent a fraction of what most models spend. Its apps are plainer than the leaders', and follow-up changes are the weak spot so far: both edits we asked for were started and not finished."},{"q":"Qwen 3.7 Plus or Qwen 3.7 Max?","a":"Plus for value on a long brief; it finished the full sign-up form. Max for a cleaner first app; it built a complete match tracker with a tidy dashboard on its first run. Both are available in Taskade Genesis."},{"q":"Can I use Qwen in Taskade?","a":"Yes. Qwen 3.7 Plus and Qwen 3.7 Max are in the Taskade model picker."},{"q":"Is Qwen open source?","a":"Qwen is an open-weight model family from Alibaba, meaning many versions publish their underlying model weights."},{"q":"Where can I compare Qwen with other models, or try it myself?","a":"Side-by-side model pages live at /compare. To try the same kind of app request yourself, start from /create. This page stays on what Qwen produced in TSK-1 tests."}]},{"slug":"grok","name":"Grok","provider":"xAI","status":"benchmarked","bestFor":"Follow-up edits that land, at a high price","verdict":"Grok gets there, slowly. Both of its August apps worked and both follow-up changes landed, which few models manage. Cost was the story: it repeated failed steps dozens of times and spent far more than the leaders on the same requests. In October, Grok 4.7 and Grok 4.6 each built a family intake form that saved every answer and both photos.","axes":{"Interface":"Strong","Task":"Strong","Memory":"Strong","Adapt":"Strong"},"score":75,"variants":[{"anchor":"grok-4-7","name":"Grok 4.7","asOf":"Oct 2026","verdict":"The newest Grok in the Taskade model picker. The 15-minute test limit ended its first complete build before a closing summary, and the family intake form it built saved every answer and both photos and took a new question cleanly.","evidence":[{"date":"2026-10-08","claim":"Built the family intake form so it saved every answer and both photos, and it refused a submission with a required answer missing."},{"date":"2026-10-08","claim":"The 15-minute test limit ended the build while it checked its own form, before a closing summary."},{"date":"2026-10-08","claim":"Added the new teacher question to the form and the teacher page, and said plainly that it had not tested a photo from a real phone."},{"date":"2026-10-08","claim":"Built the best Dota 2 match tracker of the Grok set: a three-page esports dashboard with labeled sample matches, a coach and a weekly automation. A logged match saved, and the heroes page landed. The 15-minute limit ended the build while it polished the look."}]},{"anchor":"grok-4-6","name":"Grok 4.6","asOf":"Oct 2026","verdict":"Working apps on both August tests, and both follow-up edits landed. In September it built rich sign-up data across four pages, then added a sign-in screen nobody asked for.","evidence":[{"date":"2026-08-25","claim":"Built a working match tracker with every asked piece, then ended without telling the user it was done."},{"date":"2026-08-25","claim":"Took the full 32-question sign-up form and saved every answer, but spread the data across ten projects where the leader used two."},{"date":"2026-08-25","claim":"Both follow-up edits landed: a ranked heroes page, and a text score field with next-step suggestions."},{"date":"2026-08-25","claim":"Dozens of failed build actions on each app made it one of the slowest and most expensive results of the test."},{"date":"2026-09-19","claim":"Took the full 32-question sign-up form into a 58-field table and four pages, with 31 of 32 questions word for word, then added a sign-in screen."},{"date":"2026-10-08","claim":"Built the family intake form in 14 minutes so it saved every answer and both photos, and added the new teacher question cleanly. Its teacher page showed three sample families as if they were real."}]},{"anchor":"grok-4-5","name":"Grok 4.5","asOf":"Oct 2026","verdict":"Not in the Taskade model picker. In its first test it built the most complete Grok family intake form: every answer and both photos saved, labeled sample families, a class assistant and a weekly automation.","evidence":[{"date":"2026-10-08","claim":"Built the family intake form in 7 minutes 25 seconds so it saved every answer and both photos, with labeled sample families, a class assistant and a weekly automation."},{"date":"2026-10-08","claim":"Added the new teacher question to the form and to every family card, and said plainly that it had not tested the photo picker."},{"date":"2026-10-08","claim":"Built a four-page Dota 2 match tracker with labeled sample matches that saved a logged match and added the heroes page, but its hero chart showed empty axes."}]},{"anchor":"grok-4-3","name":"Grok 4.3","asOf":"Oct 2026","verdict":"Not in the Taskade model picker. Its family intake form had hard-to-read fields and no link to the teacher page, yet it saved every answer and both photos in under 4 minutes, and the new question landed.","evidence":[{"date":"2026-10-08","claim":"Built the family intake form in 3 minutes 21 seconds so it saved every answer and both photos, but its form fields were hard to read and no link led to the teacher page."},{"date":"2026-10-08","claim":"Added the new teacher question to the form and the family cards."},{"date":"2026-10-08","claim":"Built a Dota 2 match tracker in under 5 minutes that saved a logged match and added the heroes page, but its average KDA card showed a garbled number and the board ran wider than a phone screen."}]},{"anchor":"grok-4-20","name":"Grok 4.20","asOf":"Oct 2026","verdict":"Not scored, and not in the Taskade model picker. Neither of its two October builds kept a parent's answers. A newer version can earn a fresh test.","evidence":[{"date":"2026-10-08","claim":"With reasoning on, it described the plan, stopped before writing an app, and wrote the app only on the follow-up request. That form rejected every submission and showed no message."},{"date":"2026-10-08","claim":"With reasoning off, it built a dark form where the request asked for cream paper, and the saved row lost the child's name and kept only the photo file names."},{"date":"2026-10-08","claim":"On a Dota 2 match tracker, neither setting could log a match. With reasoning on, every save was rejected and the board showed made-up stats. With reasoning off, it built no app page, and its summary described a heroes page that did not exist."}]},{"anchor":"grok-build-0-1","name":"Grok Build 0.1","asOf":"Oct 2026","verdict":"Its Dota 2 match tracker worked in October, with a logged match saved and a heroes page added. That month its family intake form dropped four of its 14 answers, one of them required. In August it built no app and repeated the same failed step hundreds of times, so October is its first test with working apps.","evidence":[{"date":"2026-08-25","claim":"Produced no projects, agents, or app: hundreds of failed build actions and no finished work."},{"date":"2026-10-08","claim":"Built the family intake form so the photos landed on every row, but 4 of 14 answers did not save, among them the required photo permission, while the form thanked the parent."},{"date":"2026-10-08","claim":"Added the new teacher question, saved it, and showed it on the teacher page."},{"date":"2026-10-08","claim":"Built a four-page Dota 2 match tracker with a win-rate trend that saved a logged match and added the heroes page when asked. Its sample matches carried no sample label."}]}],"faq":[{"q":"What is Grok best at in Taskade Genesis?","a":"Follow-up changes. Grok 4.6 was one of the few models to land both edits we asked for on its August apps, and both apps worked. The trade-off is cost and time: it repeated failed steps dozens of times and spent far more than the leaders on the same requests."},{"q":"How did Grok Build 0.1 do?","a":"In August it produced no app: it described every layer of the build and then repeated the same failed step hundreds of times until we stopped the test. In October it built a working Dota 2 match tracker, and its family intake form lost four of 14 answers. We publish the tests that go badly alongside the ones that go well."},{"q":"Can I use Grok in Taskade?","a":"Yes. Grok 4.7, Grok 4.6 and Grok Build 0.1 are in the Taskade model picker on Max and Enterprise. Grok 4.7, the newest, saved every answer and both photos on a family intake form in its first complete test."},{"q":"Where can I compare Grok with other models, or try it myself?","a":"Side-by-side model pages live at /compare. To try the same kind of app request yourself, start from /create. This page stays on what Grok produced in TSK-1 tests."}]},{"slug":"mistral","name":"Mistral","provider":"Mistral","status":"benchmarked","bestFor":"Not scored in this test","verdict":"Mistral Large 4 built a polished family intake form in its first test, but the form could never keep a photo, so no parent could submit it. We open every result before ranking it, so Mistral is not scored yet.","availability":"Mistral is not in the Taskade model picker. We tested Mistral Large 4 to compare it with the models you can pick.","axes":{"Interface":"Not scored","Task":"Not scored","Memory":"Not scored","Adapt":"Not scored"},"variants":[{"anchor":"mistral-large-4","name":"Mistral Large 4","asOf":"Oct 2026","verdict":"Not scored. Its first family intake form looked polished, but it dropped every photo a parent picked, so the two-photo rule blocked every submission. A newer version can earn a fresh test.","evidence":[{"date":"2026-10-08","claim":"Built a polished family intake form, but it dropped every photo a parent picked, so the two-photo rule blocked every submission."},{"date":"2026-10-08","claim":"The 20-minute test limit ended the build, and the follow-up turn added the new teacher question to the form and the teacher page."}]}],"faq":[{"q":"Can I use Mistral in Taskade?","a":"Not in the model picker. We tested Mistral Large 4 to compare it with the models you can pick, and TSK-1 Auto handles the default."},{"q":"Why does Mistral have no score?","a":"Its first family intake form could not take a submission: it dropped every photo a parent picked, and the form needs at least two. We open and use every result before ranking it, and a newer version can earn a fresh test."}]},{"slug":"nemotron","name":"Nemotron","provider":"NVIDIA","status":"available","bestFor":"Available in Taskade Genesis today","verdict":"Nemotron 3 Super is in the Taskade model picker. In its one August test it wrote a plan for the app but built none, so its TSK-1 score stays open until a new test.","variants":[{"anchor":"nemotron-3-super","name":"Nemotron 3 Super","asOf":"Aug 2026","verdict":"Tested once in August on the Dota 2 match tracker. It wrote out a plan but built no app in that test, so its score stays open until a new test.","evidence":[{"date":"2026-08-25","claim":"Wrote out a plan for the match tracker but created no app page, project or automation before it stopped, about 12 minutes in."}]}],"faq":[{"q":"Can I use Nemotron in Taskade?","a":"Yes. Nemotron 3 Super is in the Taskade model picker."},{"q":"Why does Nemotron have no TSK-1 score yet?","a":"Every model earns its score by building the same real app from the same request. In its one August test, Nemotron 3 Super wrote a plan for a Dota 2 match tracker but built no app, so one result is not enough to grade it. Its score stays open until a new test."}]}],"rounds":[{"anchor":"test-2026-10-08","date":"2026-10-08","title":"The family intake form, more models","request":"Parents fill in a form on their phone with two to four photos, the teacher sees every family on one page, and then we ask for one more question on the form.","results":[{"family":"grok","variant":"grok-4-7","model":"Grok 4.7","builds":1,"outcome":"Works","finding":"Saved every answer and both photos, and the new teacher question landed. The 15-minute limit ended the build before its closing summary."},{"family":"grok","variant":"grok-4-6","model":"Grok 4.6","builds":1,"outcome":"Works","finding":"Saved every answer and both photos, and the new teacher question landed. Its teacher page showed three sample families with no sample label."},{"family":"grok","variant":"grok-4-5","model":"Grok 4.5","builds":1,"outcome":"Works","finding":"Saved every answer and both photos, labeled its sample families, and added a class assistant and a weekly automation. The new teacher question landed."},{"family":"gpt","variant":"sol","model":"GPT-5.6 Sol","builds":1,"outcome":"Works","finding":"Saved every answer and both photos, labeled its sample family, and kept the teacher page open with no sign-in. The new teacher question landed."},{"family":"glm","variant":"glm-5","model":"GLM-5","builds":1,"outcome":"Works","finding":"Saved every answer and both photos, and the new teacher question showed on its gallery page, but no link led to that page."},{"family":"claude","variant":"sonnet","model":"Claude Sonnet 5","builds":2,"outcome":"Works","finding":"Both builds saved every answer and both photos, and the new teacher question landed. Sample families showed on the teacher page with no label."},{"family":"deepseek","variant":"pro-0813","model":"DeepSeek V4 Pro 0813","builds":2,"outcome":"Works","finding":"Both settings saved every answer and both photos, and the new teacher question landed. With max thinking, the follow-up turn finished the build."},{"family":"deepseek","variant":"pro-max-thinking","model":"DeepSeek V4 Pro, max thinking","builds":1,"outcome":"Works","finding":"The 15-minute limit ended the build. The follow-up turn finished a form that saved every answer and both photos, with no made-up families."},{"family":"deepseek","variant":"flash-4-1","model":"DeepSeek V4.1 Flash","builds":1,"outcome":"Works","finding":"Saved every answer and both photos, and its new-family automation ran. Three sample families showed as real."},{"family":"deepseek","variant":"pro","model":"DeepSeek V4 Pro","builds":1,"outcome":"Works","finding":"Saved every answer and both photos, and the new teacher question landed. Two sample families showed as real."},{"family":"gpt","variant":"luna","model":"GPT-5.6 Luna","builds":2,"outcome":"Works, with gaps","finding":"One build saved every answer and both photos in under 5 minutes. In the other, every parent submission failed on the photo permission question."},{"family":"gpt","variant":"terra","model":"GPT-5.6 Terra","builds":3,"outcome":"Works, with gaps","finding":"All three builds saved every answer and both photos in about 4 minutes or less. One opened in a dark look where the request asked for cream paper."},{"family":"grok","variant":"grok-4-3","model":"Grok 4.3","builds":1,"outcome":"Works, with gaps","finding":"Saved every answer and both photos in under 4 minutes, but its form fields were hard to read and no link led to the teacher page."},{"family":"claude","variant":"haiku","model":"Claude Haiku 4.5","builds":1,"outcome":"Works, with gaps","finding":"Saved every answer, but the family photo replaced the child photo on every submission."},{"family":"claude","variant":"sonnet-4-5","model":"Claude Sonnet 4.5","builds":1,"outcome":"Works, with gaps","finding":"Saved every answer and both photos in a plain look, and its open teacher page listed parent phone numbers."},{"family":"grok","variant":"grok-build-0-1","model":"Grok Build 0.1","builds":1,"outcome":"Works, with gaps","finding":"The photos landed, but 4 of 14 answers did not save, among them a required one, while the form thanked the parent."},{"family":"glm","variant":"glm-4-7","model":"GLM-4.7","builds":1,"outcome":"Works, with gaps","finding":"Saved every answer and both photos, but the teacher page crashed on load."},{"family":"deepseek","variant":"pro-high-thinking","model":"DeepSeek V4 Pro, high thinking","builds":1,"outcome":"Works, with gaps","finding":"The form dropped every photo a parent picked, so the two-photo rule blocked every submission."},{"family":"mistral","variant":"mistral-large-4","model":"Mistral Large 4","builds":1,"outcome":"Works, with gaps","finding":"A polished form, but it dropped every photo a parent picked, so the two-photo rule blocked every submission."},{"family":"deepseek","variant":"v3-2","model":"DeepSeek V3.2","builds":2,"outcome":"Works, with gaps","finding":"Neither setting saved a family: one save was rejected, and the other sent the form to an address that does not exist."},{"family":"grok","variant":"grok-4-20","model":"Grok 4.20","builds":2,"outcome":"Works, with gaps","finding":"Neither build kept a parent's answers: one lost the child's name, and the other rejected every submission with no message."}]},{"anchor":"test-2026-10-08-match-tracker","date":"2026-10-08","title":"A Dota 2 match tracker with a heroes page","request":"One player logs each match with hero, result, kills, deaths, assists, duration and notes, a dashboard shows win rate and streaks in a premium esports look, and then we ask for a heroes page.","results":[{"family":"grok","variant":"grok-4-7","model":"Grok 4.7","builds":1,"outcome":"Works","finding":"A three-page esports dashboard with labeled sample matches, a coach and a weekly automation. A logged match saved, and the heroes page landed."},{"family":"gpt","variant":"sol","model":"GPT-5.6 Sol","builds":1,"outcome":"Works","finding":"Labeled its sample matches as a demo, saved a logged match with every detail, and put the new heroes page in both menus."},{"family":"deepseek","variant":"pro-max-thinking","model":"DeepSeek V4 Pro, max thinking","builds":1,"outcome":"Works","finding":"The best-looking tracker of its set, and a logged match saved. The follow-up turn finished the interface and the heroes page."},{"family":"deepseek","variant":"pro-high-thinking","model":"DeepSeek V4 Pro, high thinking","builds":1,"outcome":"Works","finding":"Labeled its demo matches and kept them out of the stats. A logged match saved, and the heroes page landed."},{"family":"deepseek","variant":"pro-0813","model":"DeepSeek V4 Pro 0813","builds":2,"outcome":"Works","finding":"Both settings saved a logged match and ranked heroes by games. One labeled its sample matches, and one showed them as real."},{"family":"deepseek","variant":"pro","model":"DeepSeek V4 Pro","builds":1,"outcome":"Works","finding":"A win-rate trend, a logged match saved, and the heroes page landed. About 15 sample matches showed as real."},{"family":"grok","variant":"grok-4-5","model":"Grok 4.5","builds":1,"outcome":"Works","finding":"Four pages with labeled sample matches, a logged match saved, and the heroes page landed. Its hero chart showed empty axes."},{"family":"grok","variant":"grok-build-0-1","model":"Grok Build 0.1","builds":1,"outcome":"Works","finding":"Four pages with a win-rate trend, a logged match saved, and the heroes page landed. Its sample matches carried no label."},{"family":"gpt","variant":"luna","model":"GPT-5.6 Luna","builds":2,"outcome":"Works","finding":"Both builds took about 5 and a half minutes, saved a logged match with every detail, and added a ranked heroes page when asked. Sample matches showed as the player's own history."},{"family":"claude","variant":"sonnet","model":"Claude Sonnet 5","builds":2,"outcome":"Works","finding":"Both builds put a win-rate trend on the dashboard and saved a logged match. One ran wider than a phone screen, and in the other the change added a second heroes page."},{"family":"gpt","variant":"terra","model":"GPT-5.6 Terra","builds":2,"outcome":"Works","finding":"Both builds took under 4 and a half minutes and added a ranked heroes page. One labeled its starter matches, and the other ran wider than a phone screen."},{"family":"claude","variant":"haiku","model":"Claude Haiku 4.5","builds":1,"outcome":"Works, with gaps","finding":"Saved logged matches and added the heroes page, but its top-heroes chart showed empty axes."},{"family":"claude","variant":"sonnet-4-5","model":"Claude Sonnet 4.5","builds":1,"outcome":"Works, with gaps","finding":"Rejected every logged match and showed every sample match as a loss, so the win rate read 0%."},{"family":"grok","variant":"grok-4-3","model":"Grok 4.3","builds":1,"outcome":"Works, with gaps","finding":"Saved a logged match and added the heroes page, but its average KDA card showed a garbled number."},{"family":"deepseek","variant":"flash-4-1","model":"DeepSeek V4.1 Flash","builds":1,"outcome":"Works, with gaps","finding":"Five pages with labeled sample matches, but every hero showed as unknown and no match could be saved."},{"family":"grok","variant":"grok-4-20","model":"Grok 4.20","builds":2,"outcome":"Works, with gaps","finding":"Neither setting could log a match: one rejected every save, and the other built no app page while its summary described one."}]},{"anchor":"test-2026-09-30","date":"2026-09-30","title":"A family intake form on a phone","request":"Parents fill in a form on their phone with two to four photos, the teacher sees every family on one page, and then we ask for one more question on the form.","results":[{"family":"gpt","variant":"gpt-6-1-sol","model":"GPT-6.1 Sol","builds":3,"outcome":"Works","finding":"All three saved every answer and both photos, and every follow-up change landed. One teacher page misread its saved answers."},{"family":"claude","variant":"opus-5-5","model":"Claude Opus 5.5","builds":1,"outcome":"Works","finding":"Saved every answer and both photos, turned away an oversized photo, and opened in the cream look the request asked for. The follow-up change landed."},{"family":"claude","variant":"opus","model":"Claude Opus 5","builds":1,"outcome":"Works","finding":"Saved every answer and both photos, and its email automation ran. The follow-up change landed."},{"family":"claude","variant":"sonnet","model":"Claude Sonnet 5","builds":1,"outcome":"Works","finding":"Saved every answer and both photos. The follow-up change landed."},{"family":"gpt","variant":"luna","model":"GPT-5.6 Luna","builds":1,"outcome":"Works, with gaps","finding":"The most economical complete build: every answer and both photos saved. Its public assistant told a visitor a child's name."},{"family":"deepseek","variant":"flash-4-1","model":"DeepSeek V4.1 Flash","builds":1,"outcome":"Works, with gaps","finding":"Saved every answer and both photos on a two-page form, and the follow-up change landed. The build ran past the 15-minute limit."},{"family":"kimi","variant":"k3","model":"Kimi K3","builds":1,"outcome":"No app","finding":"Wrote no app page within the 15-minute limit."}]},{"anchor":"test-2026-09-22","date":"2026-09-22","title":"A match tracker","request":"One player logs each game with hero, result and notes, and a dashboard shows win rate and streaks.","results":[{"family":"gpt","variant":"luna","model":"GPT-5.6 Luna","builds":1,"outcome":"Works","finding":"The fastest and most economical complete build, in 8 minutes 19 seconds. A match saved from the app came back with all eight details."},{"family":"deepseek","variant":"flash-4-1","model":"DeepSeek V4.1 Flash","builds":1,"outcome":"Works","finding":"Three pages, and its coaching automation wrote a note into the first match logged."},{"family":"claude","variant":"opus-5-5","model":"Claude Opus 5.5","builds":1,"outcome":"Works","finding":"Asked two questions first, then built three pages and a coaching automation, at a far higher cost."},{"family":"deepseek","variant":"pro","model":"DeepSeek V4 Pro","builds":1,"outcome":"Works, with gaps","finding":"One of the two best-looking trackers of the test, but its hero picker listed only 17 heroes."},{"family":"gpt","variant":"terra","model":"GPT-5.6 Terra","builds":1,"outcome":"Works, with gaps","finding":"A clean tracker, but its match form offered only four heroes."},{"family":"gpt","variant":"gpt-6-luna","model":"GPT-6 Luna","builds":1,"outcome":"Works, with gaps","finding":"Saved a logged match and took any hero name, but the dashboard opened empty because the app hid every sample match."},{"family":"gpt","variant":"sol","model":"GPT-5.6 Sol","builds":1,"outcome":"Works, with gaps","finding":"Put a sign-in screen nobody asked for in front of a personal tracker."}]}],"tasks":[{"slug":"intake-form","arena":"build","name":"intake form","summary":"Parents fill in a form on their phone, with photos, and a teacher sees every family on one page. Every required answer and every photo must save.","pages":[{"arena":"build","url":"https://www.taskade.com/tsk/build/intake-form","requests":[{"date":"2026-10-08","request":"Parents fill in a form on their phone with two to four photos, the teacher sees every family on one page, and then we ask for one more question on the form."}],"results":[{"rank":1,"family":"claude","variant":"opus","model":"Claude Opus 5","builds":1,"outcome":"Works","finding":"Saved every answer and both photos, and its email automation ran. The follow-up change landed.","date":"2026-09-30","test":"test-2026-09-30"},{"rank":1,"family":"claude","variant":"opus-5-5","model":"Claude Opus 5.5","builds":1,"outcome":"Works","finding":"Saved every answer and both photos, turned away an oversized photo, and opened in the cream look the request asked for. The follow-up change landed.","date":"2026-09-30","test":"test-2026-09-30"},{"rank":1,"family":"claude","variant":"sonnet","model":"Claude Sonnet 5","builds":2,"outcome":"Works","finding":"Both builds saved every answer and both photos, and the new teacher question landed. Sample families showed on the teacher page with no label.","date":"2026-10-08","test":"test-2026-10-08"},{"rank":1,"family":"deepseek","variant":"pro","model":"DeepSeek V4 Pro","builds":1,"outcome":"Works","finding":"Saved every answer and both photos, and the new teacher question landed. Two sample families showed as real.","date":"2026-10-08","test":"test-2026-10-08"},{"rank":1,"family":"deepseek","variant":"pro-0813","model":"DeepSeek V4 Pro 0813","builds":2,"outcome":"Works","finding":"Both settings saved every answer and both photos, and the new teacher question landed. With max thinking, the follow-up turn finished the build.","date":"2026-10-08","test":"test-2026-10-08"},{"rank":1,"family":"deepseek","variant":"pro-max-thinking","model":"DeepSeek V4 Pro, max thinking","builds":1,"outcome":"Works","finding":"The 15-minute limit ended the build. The follow-up turn finished a form that saved every answer and both photos, with no made-up families.","date":"2026-10-08","test":"test-2026-10-08"},{"rank":1,"family":"deepseek","variant":"flash-4-1","model":"DeepSeek V4.1 Flash","builds":1,"outcome":"Works","finding":"Saved every answer and both photos, and its new-family automation ran. Three sample families showed as real.","date":"2026-10-08","test":"test-2026-10-08"},{"rank":1,"family":"glm","variant":"glm-5","model":"GLM-5","builds":1,"outcome":"Works","finding":"Saved every answer and both photos, and the new teacher question showed on its gallery page, but no link led to that page.","date":"2026-10-08","test":"test-2026-10-08"},{"rank":1,"family":"gpt","variant":"sol","model":"GPT-5.6 Sol","builds":1,"outcome":"Works","finding":"Saved every answer and both photos, labeled its sample family, and kept the teacher page open with no sign-in. The new teacher question landed.","date":"2026-10-08","test":"test-2026-10-08"},{"rank":1,"family":"gpt","variant":"gpt-6-1-sol","model":"GPT-6.1 Sol","builds":3,"outcome":"Works","finding":"All three saved every answer and both photos, and every follow-up change landed. One teacher page misread its saved answers.","date":"2026-09-30","test":"test-2026-09-30"},{"rank":1,"family":"grok","variant":"grok-4-5","model":"Grok 4.5","builds":1,"outcome":"Works","finding":"Saved every answer and both photos, labeled its sample families, and added a class assistant and a weekly automation. The new teacher question landed.","date":"2026-10-08","test":"test-2026-10-08"},{"rank":1,"family":"grok","variant":"grok-4-6","model":"Grok 4.6","builds":1,"outcome":"Works","finding":"Saved every answer and both photos, and the new teacher question landed. Its teacher page showed three sample families with no sample label.","date":"2026-10-08","test":"test-2026-10-08"},{"rank":1,"family":"grok","variant":"grok-4-7","model":"Grok 4.7","builds":1,"outcome":"Works","finding":"Saved every answer and both photos, and the new teacher question landed. The 15-minute limit ended the build before its closing summary.","date":"2026-10-08","test":"test-2026-10-08"},{"rank":14,"family":"claude","variant":"haiku","model":"Claude Haiku 4.5","builds":1,"outcome":"Works, with gaps","finding":"Saved every answer, but the family photo replaced the child photo on every submission.","date":"2026-10-08","test":"test-2026-10-08"},{"rank":14,"family":"claude","variant":"sonnet-4-5","model":"Claude Sonnet 4.5","builds":1,"outcome":"Works, with gaps","finding":"Saved every answer and both photos in a plain look, and its open teacher page listed parent phone numbers.","date":"2026-10-08","test":"test-2026-10-08"},{"rank":14,"family":"deepseek","variant":"v3-2","model":"DeepSeek V3.2","builds":2,"outcome":"Works, with gaps","finding":"Neither setting saved a family: one save was rejected, and the other sent the form to an address that does not exist.","date":"2026-10-08","test":"test-2026-10-08"},{"rank":14,"family":"deepseek","variant":"pro-high-thinking","model":"DeepSeek V4 Pro, high thinking","builds":1,"outcome":"Works, with gaps","finding":"The form dropped every photo a parent picked, so the two-photo rule blocked every submission.","date":"2026-10-08","test":"test-2026-10-08"},{"rank":14,"family":"glm","variant":"glm-4-7","model":"GLM-4.7","builds":1,"outcome":"Works, with gaps","finding":"Saved every answer and both photos, but the teacher page crashed on load.","date":"2026-10-08","test":"test-2026-10-08"},{"rank":14,"family":"gpt","variant":"luna","model":"GPT-5.6 Luna","builds":2,"outcome":"Works, with gaps","finding":"One build saved every answer and both photos in under 5 minutes. In the other, every parent submission failed on the photo permission question.","date":"2026-10-08","test":"test-2026-10-08"},{"rank":14,"family":"gpt","variant":"terra","model":"GPT-5.6 Terra","builds":3,"outcome":"Works, with gaps","finding":"All three builds saved every answer and both photos in about 4 minutes or less. One opened in a dark look where the request asked for cream paper.","date":"2026-10-08","test":"test-2026-10-08"},{"rank":14,"family":"grok","variant":"grok-4-3","model":"Grok 4.3","builds":1,"outcome":"Works, with gaps","finding":"Saved every answer and both photos in under 4 minutes, but its form fields were hard to read and no link led to the teacher page.","date":"2026-10-08","test":"test-2026-10-08"},{"rank":14,"family":"grok","variant":"grok-4-20","model":"Grok 4.20","builds":2,"outcome":"Works, with gaps","finding":"Neither build kept a parent's answers: one lost the child's name, and the other rejected every submission with no message.","date":"2026-10-08","test":"test-2026-10-08"},{"rank":14,"family":"grok","variant":"grok-build-0-1","model":"Grok Build 0.1","builds":1,"outcome":"Works, with gaps","finding":"The photos landed, but 4 of 14 answers did not save, among them a required one, while the form thanked the parent.","date":"2026-10-08","test":"test-2026-10-08"},{"rank":14,"family":"mistral","variant":"mistral-large-4","model":"Mistral Large 4","builds":1,"outcome":"Works, with gaps","finding":"A polished form, but it dropped every photo a parent picked, so the two-photo rule blocked every submission.","date":"2026-10-08","test":"test-2026-10-08"},{"rank":25,"family":"kimi","variant":"k3","model":"Kimi K3","builds":1,"outcome":"No app","finding":"Wrote no app page within the 15-minute limit.","date":"2026-09-30","test":"test-2026-09-30"}]}]},{"slug":"game-tracker","arena":"build","name":"game tracker app","summary":"One Dota 2 player logs every match with the hero, the result and the stats. A dashboard shows win rate and streaks.","pages":[{"arena":"build","url":"https://www.taskade.com/tsk/build/game-tracker","requests":[{"date":"2026-10-08","request":"One player logs each match with hero, result, kills, deaths, assists, duration and notes, a dashboard shows win rate and streaks in a premium esports look, and then we ask for a heroes page."}],"results":[{"rank":1,"family":"claude","variant":"sonnet","model":"Claude Sonnet 5","builds":2,"outcome":"Works","finding":"Both builds put a win-rate trend on the dashboard and saved a logged match. One ran wider than a phone screen, and in the other the change added a second heroes page.","date":"2026-10-08","test":"test-2026-10-08-match-tracker"},{"rank":1,"family":"deepseek","variant":"pro","model":"DeepSeek V4 Pro","builds":1,"outcome":"Works","finding":"A win-rate trend, a logged match saved, and the heroes page landed. About 15 sample matches showed as real.","date":"2026-10-08","test":"test-2026-10-08-match-tracker"},{"rank":1,"family":"deepseek","variant":"pro-0813","model":"DeepSeek V4 Pro 0813","builds":2,"outcome":"Works","finding":"Both settings saved a logged match and ranked heroes by games. One labeled its sample matches, and one showed them as real.","date":"2026-10-08","test":"test-2026-10-08-match-tracker"},{"rank":1,"family":"deepseek","variant":"pro-high-thinking","model":"DeepSeek V4 Pro, high thinking","builds":1,"outcome":"Works","finding":"Labeled its demo matches and kept them out of the stats. A logged match saved, and the heroes page landed.","date":"2026-10-08","test":"test-2026-10-08-match-tracker"},{"rank":1,"family":"deepseek","variant":"pro-max-thinking","model":"DeepSeek V4 Pro, max thinking","builds":1,"outcome":"Works","finding":"The best-looking tracker of its set, and a logged match saved. The follow-up turn finished the interface and the heroes page.","date":"2026-10-08","test":"test-2026-10-08-match-tracker"},{"rank":1,"family":"gpt","variant":"luna","model":"GPT-5.6 Luna","builds":2,"outcome":"Works","finding":"Both builds took about 5 and a half minutes, saved a logged match with every detail, and added a ranked heroes page when asked. Sample matches showed as the player's own history.","date":"2026-10-08","test":"test-2026-10-08-match-tracker"},{"rank":1,"family":"gpt","variant":"sol","model":"GPT-5.6 Sol","builds":1,"outcome":"Works","finding":"Labeled its sample matches as a demo, saved a logged match with every detail, and put the new heroes page in both menus.","date":"2026-10-08","test":"test-2026-10-08-match-tracker"},{"rank":1,"family":"gpt","variant":"terra","model":"GPT-5.6 Terra","builds":2,"outcome":"Works","finding":"Both builds took under 4 and a half minutes and added a ranked heroes page. One labeled its starter matches, and the other ran wider than a phone screen.","date":"2026-10-08","test":"test-2026-10-08-match-tracker"},{"rank":1,"family":"grok","variant":"grok-4-5","model":"Grok 4.5","builds":1,"outcome":"Works","finding":"Four pages with labeled sample matches, a logged match saved, and the heroes page landed. Its hero chart showed empty axes.","date":"2026-10-08","test":"test-2026-10-08-match-tracker"},{"rank":1,"family":"grok","variant":"grok-4-7","model":"Grok 4.7","builds":1,"outcome":"Works","finding":"A three-page esports dashboard with labeled sample matches, a coach and a weekly automation. A logged match saved, and the heroes page landed.","date":"2026-10-08","test":"test-2026-10-08-match-tracker"},{"rank":1,"family":"grok","variant":"grok-build-0-1","model":"Grok Build 0.1","builds":1,"outcome":"Works","finding":"Four pages with a win-rate trend, a logged match saved, and the heroes page landed. Its sample matches carried no label.","date":"2026-10-08","test":"test-2026-10-08-match-tracker"},{"rank":12,"family":"claude","variant":"haiku","model":"Claude Haiku 4.5","builds":1,"outcome":"Works, with gaps","finding":"Saved logged matches and added the heroes page, but its top-heroes chart showed empty axes.","date":"2026-10-08","test":"test-2026-10-08-match-tracker"},{"rank":12,"family":"claude","variant":"sonnet-4-5","model":"Claude Sonnet 4.5","builds":1,"outcome":"Works, with gaps","finding":"Rejected every logged match and showed every sample match as a loss, so the win rate read 0%.","date":"2026-10-08","test":"test-2026-10-08-match-tracker"},{"rank":12,"family":"deepseek","variant":"flash-4-1","model":"DeepSeek V4.1 Flash","builds":1,"outcome":"Works, with gaps","finding":"Five pages with labeled sample matches, but every hero showed as unknown and no match could be saved.","date":"2026-10-08","test":"test-2026-10-08-match-tracker"},{"rank":12,"family":"grok","variant":"grok-4-3","model":"Grok 4.3","builds":1,"outcome":"Works, with gaps","finding":"Saved a logged match and added the heroes page, but its average KDA card showed a garbled number.","date":"2026-10-08","test":"test-2026-10-08-match-tracker"},{"rank":12,"family":"grok","variant":"grok-4-20","model":"Grok 4.20","builds":2,"outcome":"Works, with gaps","finding":"Neither setting could log a match: one rejected every save, and the other built no app page while its summary described one.","date":"2026-10-08","test":"test-2026-10-08-match-tracker"}]}]},{"slug":"crm","arena":"build","name":"sales CRM","summary":"A sales team captures leads, and an AI agent scores each one by fit. A board sorts the leads from hot to cold, and a daily digest lists the best ones.","pages":[{"arena":"build","url":"https://www.taskade.com/tsk/build/crm","requests":[],"results":[]}]},{"slug":"inspection","arena":"build","name":"inspection app","summary":"Inspectors run a weekly pass or fail checklist on each van. A dashboard shows what failed, and a reminder goes out when a check is overdue.","pages":[{"arena":"build","url":"https://www.taskade.com/tsk/build/inspection","requests":[],"results":[]}]},{"slug":"governance","arena":"build","name":"AI governance tracker","summary":"Teams ask for approval before they use an AI tool, and a reviewer approves or rejects each request. Every run goes in a log, and old records become read-only.","pages":[{"arena":"build","url":"https://www.taskade.com/tsk/build/governance","requests":[],"results":[]}]},{"slug":"reading-list","arena":"build","name":"reading list app","summary":"Save books with the title, the author and a status. Filter the list by to read, reading or done.","pages":[{"arena":"build","url":"https://www.taskade.com/tsk/build/reading-list","requests":[],"results":[]}]},{"slug":"habit-tracker","arena":"build","name":"habit tracker app","summary":"Check off daily habits and keep a streak for each one. A dashboard shows progress across every habit.","pages":[{"arena":"build","url":"https://www.taskade.com/tsk/build/habit-tracker","requests":[],"results":[]}]},{"slug":"recipes","arena":"build","name":"recipe app","summary":"Save recipes with ingredients, steps and tags. Browse the cookbook and print any recipe.","pages":[{"arena":"build","url":"https://www.taskade.com/tsk/build/recipes","requests":[],"results":[]}]},{"slug":"new-page","arena":"code","name":"Add a page to an app you built","summary":"Add a new page to an app the model already built. The new page must work, and every page that worked before must still work.","app":"game-tracker","followUpRequest":"Add a heroes page.","pages":[{"arena":"code","url":"https://www.taskade.com/tsk/code/new-page","requests":[{"date":"2026-10-08","request":"One player logs each match with hero, result, kills, deaths, assists, duration and notes, a dashboard shows win rate and streaks in a premium esports look, and then we ask for a heroes page."}],"results":[{"rank":1,"family":"claude","variant":"sonnet","model":"Claude Sonnet 5","builds":2,"outcome":"Works","finding":"Both builds put a win-rate trend on the dashboard and saved a logged match. One ran wider than a phone screen, and in the other the change added a second heroes page.","date":"2026-10-08","test":"test-2026-10-08-match-tracker"},{"rank":1,"family":"deepseek","variant":"pro","model":"DeepSeek V4 Pro","builds":1,"outcome":"Works","finding":"A win-rate trend, a logged match saved, and the heroes page landed. About 15 sample matches showed as real.","date":"2026-10-08","test":"test-2026-10-08-match-tracker"},{"rank":1,"family":"deepseek","variant":"pro-0813","model":"DeepSeek V4 Pro 0813","builds":2,"outcome":"Works","finding":"Both settings saved a logged match and ranked heroes by games. One labeled its sample matches, and one showed them as real.","date":"2026-10-08","test":"test-2026-10-08-match-tracker"},{"rank":1,"family":"deepseek","variant":"pro-high-thinking","model":"DeepSeek V4 Pro, high thinking","builds":1,"outcome":"Works","finding":"Labeled its demo matches and kept them out of the stats. A logged match saved, and the heroes page landed.","date":"2026-10-08","test":"test-2026-10-08-match-tracker"},{"rank":1,"family":"deepseek","variant":"pro-max-thinking","model":"DeepSeek V4 Pro, max thinking","builds":1,"outcome":"Works","finding":"The best-looking tracker of its set, and a logged match saved. The follow-up turn finished the interface and the heroes page.","date":"2026-10-08","test":"test-2026-10-08-match-tracker"},{"rank":1,"family":"gpt","variant":"luna","model":"GPT-5.6 Luna","builds":2,"outcome":"Works","finding":"Both builds took about 5 and a half minutes, saved a logged match with every detail, and added a ranked heroes page when asked. Sample matches showed as the player's own history.","date":"2026-10-08","test":"test-2026-10-08-match-tracker"},{"rank":1,"family":"gpt","variant":"sol","model":"GPT-5.6 Sol","builds":1,"outcome":"Works","finding":"Labeled its sample matches as a demo, saved a logged match with every detail, and put the new heroes page in both menus.","date":"2026-10-08","test":"test-2026-10-08-match-tracker"},{"rank":1,"family":"gpt","variant":"terra","model":"GPT-5.6 Terra","builds":2,"outcome":"Works","finding":"Both builds took under 4 and a half minutes and added a ranked heroes page. One labeled its starter matches, and the other ran wider than a phone screen.","date":"2026-10-08","test":"test-2026-10-08-match-tracker"},{"rank":1,"family":"grok","variant":"grok-4-5","model":"Grok 4.5","builds":1,"outcome":"Works","finding":"Four pages with labeled sample matches, a logged match saved, and the heroes page landed. Its hero chart showed empty axes.","date":"2026-10-08","test":"test-2026-10-08-match-tracker"},{"rank":1,"family":"grok","variant":"grok-4-7","model":"Grok 4.7","builds":1,"outcome":"Works","finding":"A three-page esports dashboard with labeled sample matches, a coach and a weekly automation. A logged match saved, and the heroes page landed.","date":"2026-10-08","test":"test-2026-10-08-match-tracker"},{"rank":1,"family":"grok","variant":"grok-build-0-1","model":"Grok Build 0.1","builds":1,"outcome":"Works","finding":"Four pages with a win-rate trend, a logged match saved, and the heroes page landed. Its sample matches carried no label.","date":"2026-10-08","test":"test-2026-10-08-match-tracker"},{"rank":12,"family":"claude","variant":"haiku","model":"Claude Haiku 4.5","builds":1,"outcome":"Works, with gaps","finding":"Saved logged matches and added the heroes page, but its top-heroes chart showed empty axes.","date":"2026-10-08","test":"test-2026-10-08-match-tracker"},{"rank":12,"family":"claude","variant":"sonnet-4-5","model":"Claude Sonnet 4.5","builds":1,"outcome":"Works, with gaps","finding":"Rejected every logged match and showed every sample match as a loss, so the win rate read 0%.","date":"2026-10-08","test":"test-2026-10-08-match-tracker"},{"rank":12,"family":"deepseek","variant":"flash-4-1","model":"DeepSeek V4.1 Flash","builds":1,"outcome":"Works, with gaps","finding":"Five pages with labeled sample matches, but every hero showed as unknown and no match could be saved.","date":"2026-10-08","test":"test-2026-10-08-match-tracker"},{"rank":12,"family":"grok","variant":"grok-4-3","model":"Grok 4.3","builds":1,"outcome":"Works, with gaps","finding":"Saved a logged match and added the heroes page, but its average KDA card showed a garbled number.","date":"2026-10-08","test":"test-2026-10-08-match-tracker"},{"rank":12,"family":"grok","variant":"grok-4-20","model":"Grok 4.20","builds":2,"outcome":"Works, with gaps","finding":"Neither setting could log a match: one rejected every save, and the other built no app page while its summary described one.","date":"2026-10-08","test":"test-2026-10-08-match-tracker"}]}]},{"slug":"new-field","arena":"code","name":"Add a field to a form","summary":"Add one more question to a form the model already built. The new answer must save, and every earlier answer must still save.","app":"intake-form","followUpRequest":"Add one more question to the form.","pages":[{"arena":"code","url":"https://www.taskade.com/tsk/code/new-field","requests":[{"date":"2026-10-08","request":"Parents fill in a form on their phone with two to four photos, the teacher sees every family on one page, and then we ask for one more question on the form."}],"results":[{"rank":1,"family":"claude","variant":"opus","model":"Claude Opus 5","builds":1,"outcome":"Works","finding":"Saved every answer and both photos, and its email automation ran. The follow-up change landed.","date":"2026-09-30","test":"test-2026-09-30"},{"rank":1,"family":"claude","variant":"opus-5-5","model":"Claude Opus 5.5","builds":1,"outcome":"Works","finding":"Saved every answer and both photos, turned away an oversized photo, and opened in the cream look the request asked for. The follow-up change landed.","date":"2026-09-30","test":"test-2026-09-30"},{"rank":1,"family":"claude","variant":"sonnet","model":"Claude Sonnet 5","builds":2,"outcome":"Works","finding":"Both builds saved every answer and both photos, and the new teacher question landed. Sample families showed on the teacher page with no label.","date":"2026-10-08","test":"test-2026-10-08"},{"rank":1,"family":"deepseek","variant":"pro","model":"DeepSeek V4 Pro","builds":1,"outcome":"Works","finding":"Saved every answer and both photos, and the new teacher question landed. Two sample families showed as real.","date":"2026-10-08","test":"test-2026-10-08"},{"rank":1,"family":"deepseek","variant":"pro-0813","model":"DeepSeek V4 Pro 0813","builds":2,"outcome":"Works","finding":"Both settings saved every answer and both photos, and the new teacher question landed. With max thinking, the follow-up turn finished the build.","date":"2026-10-08","test":"test-2026-10-08"},{"rank":1,"family":"deepseek","variant":"pro-max-thinking","model":"DeepSeek V4 Pro, max thinking","builds":1,"outcome":"Works","finding":"The 15-minute limit ended the build. The follow-up turn finished a form that saved every answer and both photos, with no made-up families.","date":"2026-10-08","test":"test-2026-10-08"},{"rank":1,"family":"deepseek","variant":"flash-4-1","model":"DeepSeek V4.1 Flash","builds":1,"outcome":"Works","finding":"Saved every answer and both photos, and its new-family automation ran. Three sample families showed as real.","date":"2026-10-08","test":"test-2026-10-08"},{"rank":1,"family":"glm","variant":"glm-5","model":"GLM-5","builds":1,"outcome":"Works","finding":"Saved every answer and both photos, and the new teacher question showed on its gallery page, but no link led to that page.","date":"2026-10-08","test":"test-2026-10-08"},{"rank":1,"family":"gpt","variant":"sol","model":"GPT-5.6 Sol","builds":1,"outcome":"Works","finding":"Saved every answer and both photos, labeled its sample family, and kept the teacher page open with no sign-in. The new teacher question landed.","date":"2026-10-08","test":"test-2026-10-08"},{"rank":1,"family":"gpt","variant":"gpt-6-1-sol","model":"GPT-6.1 Sol","builds":3,"outcome":"Works","finding":"All three saved every answer and both photos, and every follow-up change landed. One teacher page misread its saved answers.","date":"2026-09-30","test":"test-2026-09-30"},{"rank":1,"family":"grok","variant":"grok-4-5","model":"Grok 4.5","builds":1,"outcome":"Works","finding":"Saved every answer and both photos, labeled its sample families, and added a class assistant and a weekly automation. The new teacher question landed.","date":"2026-10-08","test":"test-2026-10-08"},{"rank":1,"family":"grok","variant":"grok-4-6","model":"Grok 4.6","builds":1,"outcome":"Works","finding":"Saved every answer and both photos, and the new teacher question landed. Its teacher page showed three sample families with no sample label.","date":"2026-10-08","test":"test-2026-10-08"},{"rank":1,"family":"grok","variant":"grok-4-7","model":"Grok 4.7","builds":1,"outcome":"Works","finding":"Saved every answer and both photos, and the new teacher question landed. The 15-minute limit ended the build before its closing summary.","date":"2026-10-08","test":"test-2026-10-08"},{"rank":14,"family":"claude","variant":"haiku","model":"Claude Haiku 4.5","builds":1,"outcome":"Works, with gaps","finding":"Saved every answer, but the family photo replaced the child photo on every submission.","date":"2026-10-08","test":"test-2026-10-08"},{"rank":14,"family":"claude","variant":"sonnet-4-5","model":"Claude Sonnet 4.5","builds":1,"outcome":"Works, with gaps","finding":"Saved every answer and both photos in a plain look, and its open teacher page listed parent phone numbers.","date":"2026-10-08","test":"test-2026-10-08"},{"rank":14,"family":"deepseek","variant":"v3-2","model":"DeepSeek V3.2","builds":2,"outcome":"Works, with gaps","finding":"Neither setting saved a family: one save was rejected, and the other sent the form to an address that does not exist.","date":"2026-10-08","test":"test-2026-10-08"},{"rank":14,"family":"deepseek","variant":"pro-high-thinking","model":"DeepSeek V4 Pro, high thinking","builds":1,"outcome":"Works, with gaps","finding":"The form dropped every photo a parent picked, so the two-photo rule blocked every submission.","date":"2026-10-08","test":"test-2026-10-08"},{"rank":14,"family":"glm","variant":"glm-4-7","model":"GLM-4.7","builds":1,"outcome":"Works, with gaps","finding":"Saved every answer and both photos, but the teacher page crashed on load.","date":"2026-10-08","test":"test-2026-10-08"},{"rank":14,"family":"gpt","variant":"luna","model":"GPT-5.6 Luna","builds":2,"outcome":"Works, with gaps","finding":"One build saved every answer and both photos in under 5 minutes. In the other, every parent submission failed on the photo permission question.","date":"2026-10-08","test":"test-2026-10-08"},{"rank":14,"family":"gpt","variant":"terra","model":"GPT-5.6 Terra","builds":3,"outcome":"Works, with gaps","finding":"All three builds saved every answer and both photos in about 4 minutes or less. One opened in a dark look where the request asked for cream paper.","date":"2026-10-08","test":"test-2026-10-08"},{"rank":14,"family":"grok","variant":"grok-4-3","model":"Grok 4.3","builds":1,"outcome":"Works, with gaps","finding":"Saved every answer and both photos in under 4 minutes, but its form fields were hard to read and no link led to the teacher page.","date":"2026-10-08","test":"test-2026-10-08"},{"rank":14,"family":"grok","variant":"grok-4-20","model":"Grok 4.20","builds":2,"outcome":"Works, with gaps","finding":"Neither build kept a parent's answers: one lost the child's name, and the other rejected every submission with no message.","date":"2026-10-08","test":"test-2026-10-08"},{"rank":14,"family":"grok","variant":"grok-build-0-1","model":"Grok Build 0.1","builds":1,"outcome":"Works, with gaps","finding":"The photos landed, but 4 of 14 answers did not save, among them a required one, while the form thanked the parent.","date":"2026-10-08","test":"test-2026-10-08"},{"rank":14,"family":"mistral","variant":"mistral-large-4","model":"Mistral Large 4","builds":1,"outcome":"Works, with gaps","finding":"A polished form, but it dropped every photo a parent picked, so the two-photo rule blocked every submission.","date":"2026-10-08","test":"test-2026-10-08"},{"rank":25,"family":"kimi","variant":"k3","model":"Kimi K3","builds":1,"outcome":"No app","finding":"Wrote no app page within the 15-minute limit.","date":"2026-09-30","test":"test-2026-09-30"}]}]},{"slug":"edit-copy","arena":"code","name":"Change text and headings","summary":"Change a heading or a line of text in an app the model already built. Only the words change, and nothing else breaks.","app":"reading-list","followUpRequest":"Change the page heading.","pages":[{"arena":"code","url":"https://www.taskade.com/tsk/code/edit-copy","requests":[],"results":[]}]}],"resultDays":[{"anchor":"results-2026-10-10","date":"2026-10-10","headline":"GPT-6 Luna on the match tracker","summary":"GPT-6 Luna built the Dota 2 match tracker with every detail the request named, a coach agent and a weekly recap, then added a heroes page in a follow-up change and put it live. One build, so read it as directional."},{"anchor":"results-2026-10-08","date":"2026-10-08","headline":"More models on two real apps","summary":"More than 20 models and settings rebuilt the family intake form and a Dota 2 match tracker, and then took one follow-up change each. Grok 4.7 and GPT-5.6 Sol built both apps well, several builds showed sample families or matches as if they were real, and a few forms could not save at all."},{"anchor":"results-2026-09-30","date":"2026-09-30","headline":"A phone form with photos","summary":"A real customer's family intake form joined the test set: parents fill it in on a phone with two to four photos, and the teacher sees every family on one page. GPT-6.1 Sol and Claude Opus 5.5 shared the top result, and every model that finished also added a follow-up question cleanly."},{"anchor":"results-2026-09-22","date":"2026-09-22","headline":"One match tracker, model by model","summary":"Several models built the same match tracker. GPT-5.6 Luna was the fastest and most economical complete build, DeepSeek V4.1 Flash ran a coaching automation on the first match logged, and two models limited their match forms to a short list of heroes."},{"anchor":"results-2026-09-19","date":"2026-09-19","headline":"Real-customer shapes and follow-up edits","summary":"Three real-customer app shapes joined the set: a sales CRM, a fleet inspection, and an AI governance office. The first full follow-up-edit ladder passed on every model that ran it."},{"anchor":"results-2026-09-18","date":"2026-09-18","headline":"A broad model sweep and parallel helpers","summary":"Seventeen models built the same client sign-up form. A separate test of parallel helpers showed that one file per helper produced the cleanest result."},{"anchor":"results-2026-08-29","date":"2026-08-29","headline":"Same request, four setups","summary":"A control run told us how big a difference has to be before it counts. Two identical setups produced noticeably different costs for the same app, so we now only claim a difference when it is larger than that gap, and we grade on what the app does."},{"anchor":"results-2026-08-28","date":"2026-08-28","headline":"Same result after a platform update","summary":"We re-ran the automatic model after a large platform update. The apps came out the same, with fewer failed build actions along the way."},{"anchor":"results-2026-08-27","date":"2026-08-27","headline":"Sixteen builds, five models, identical prompts","summary":"Five models built and edited the same apps from identical prompts. GPT-5.6 Luna was the value pick on every app and GPT-5.6 Terra the quality pick. Most builds now keep a checklist of the request inside the app."},{"anchor":"results-2026-08-26","date":"2026-08-26","headline":"Cleaner follow-up edits","summary":"Follow-up edits now start from a map of the app, and they landed cleaner. GPT-5.6 Luna edited apps other models had built without breaking them, and a missing page now explains itself."},{"anchor":"results-2026-08-25","date":"2026-08-25","headline":"Two new model families, three new app shapes","summary":"Qwen and Grok entered the benchmark, and three real-customer app shapes joined the test set: a sales CRM, a field inspection audit, and an AI governance office. Every model still gets the same words."},{"anchor":"results-2026-08-20","date":"2026-08-20","headline":"A dedicated follow-up-edit test","summary":"We added a dedicated follow-up-edit pass and stricter instruction checks. On its first run, none of the 31 edit attempts finished - a strong first build is not enough on its own; an app must also change cleanly and stay within the brief."},{"anchor":"results-2026-08-19","date":"2026-08-19","headline":"Faster builds, verified results","summary":"A complete app finished in under five minutes without follow-up. We also rechecked every published result and established a baseline for automatic model selection."},{"anchor":"results-2026-08-08","date":"2026-08-08","headline":"Accuracy and value moved together","summary":"The strongest results reproduced a long brief exactly, balanced build quality with efficiency, and explained their design choices clearly."},{"anchor":"results-2026-08-07","date":"2026-08-07","headline":"Lower cost did not mean lower quality","summary":"An efficient build completed a long form at the lowest measured cost, while the leading result also saved every submitted field correctly."},{"anchor":"results-2026-08-06","date":"2026-08-06","headline":"Long briefs became a harder test","summary":"Several builds followed the brief word for word, and one handled a 60-field form without cutting requirements."},{"anchor":"results-2026-08-05","date":"2026-08-05","headline":"Working software became the baseline","summary":"We made a running app the entry requirement. Attractive results that failed to open no longer received a score."},{"anchor":"results-2026-08-04","date":"2026-08-04","headline":"Efficiency without tradeoffs","summary":"The cleanest test also delivered the lowest measured cost, showing that careful model selection can improve both quality and value."},{"anchor":"results-2026-08-03","date":"2026-08-03","headline":"Brief fidelity and saved data","summary":"The strongest builds matched every requested field, stored submitted data correctly, and remained useful from form to follow-up."},{"anchor":"results-2026-08-02","date":"2026-08-02","headline":"Prompt length changed the outcome","summary":"Shorter instructions improved brief matching and edits, but sometimes reduced visual quality. The benchmark now balances all four dimensions."},{"anchor":"results-2026-08-01","date":"2026-08-01","headline":"Nine models, one consistent standard","summary":"Nine models built the same app. Several produced polished designs, but only working results qualified for comparison."},{"anchor":"results-2026-07-31","date":"2026-07-31","headline":"Self-checking improved reliability","summary":"The most dependable results detected and repaired problems before finishing, while weaker builds reported success too early."},{"anchor":"results-2026-07-30","date":"2026-07-30","headline":"Restraint mattered","summary":"The strongest designs followed the brief without adding unwanted gates or steps. Doing only what was asked became part of quality."}],"meta":{"lastUpdated":"2026-10-10","faq":[{"q":"Which AI model is best for building apps?","a":"Some models hold closer to a brief, some finish faster, some look more polished. In Taskade you do not have to choose: TSK-1 Auto handles the default, and you can pick a model yourself when you want more control."},{"q":"What can Taskade Genesis build with TSK-1?","a":"Taskade Genesis can turn one request into CRMs, dashboards, portals, forms, trackers, and internal tools. TSK-1 coordinates the AI model, workspace memory, agents, and automations so the result can store information, answer questions, run workflows, and keep improving."},{"q":"What does the TSK-1 Benchmark measure?","a":"Whether a model can turn one prompt into a complete, working app. We grade four qualities of a living system: Interface, Task, Memory, and Adaptation: how finished it feels, how closely it follows your request, whether it keeps your data, and how cleanly it handles follow-up changes."},{"q":"How is TSK-1 different from Vibe Code Bench, WebDev Arena, and Design Arena?","a":"Those benchmarks grade code, a web page, or crowd votes on a screenshot. TSK-1 grades a running app: we open it, enter data the way a customer would, check that the data was kept, run its automation, and ask for a follow-up change. A build that looks right but loses your data does not score."},{"q":"Does TSK-1 benchmark app builders or AI models?","a":"AI models. Every model builds inside the same app builder, Taskade Genesis, from the same request, word for word, so the builder is held constant. A control run in August showed that two identical setups can still differ, so we only call out a difference between models when it is larger than that gap."},{"q":"Do you test the apps or just the code?","a":"We test the apps. Every build gets opened and used. We fill in its form the way a customer would, submit it, then check that the answers landed in the right place. An app that looks beautiful and loses your data fails."},{"q":"Which models can I use in Taskade?","a":"Taskade runs frontier models from top AI labs, and TSK-1 Auto handles the default. Built in, with no key, every plan can also pick the TSK-Instant rung. The AI model directory lists every model, with its status, its plan and how to pick it."},{"q":"How often is this updated?","a":"We run a new test whenever a notable model ships. Every result cleared for public comparison stays on this page, including the ones that did not go well."},{"q":"What is TSK-1?","a":"TSK-1 is the intelligence behind Taskade Genesis. It brings AI models, workspace memory, agents, and automations together so one request can become a working app. The TSK-1 Benchmark shows how that process performs on real app builds."},{"q":"Can I try the same app request?","a":"Yes. Every model in a comparison receives the same request, word for word. Describe the same idea in Taskade Genesis, choose a model or let TSK-1 Auto decide, and see what it builds."},{"q":"Are the benchmarked apps publicly available?","a":"Not yet one by one. We are listing every finished benchmark build under a single official creator so you can open and clone it. Until then, each model page explains what that model produced, and the App Kits on this page are live systems you can open and clone today."},{"q":"How many builds are behind each result?","a":"Each result comes from a small set of hands-on app builds. The TSK Score rescales four evidence tiers to 0-100 for easier comparison; it is not 100 separate checks. Models with the same result share a rank."}]}}