TSK-Bench · Code arena
Which AI model adds a page to an app without breaking it?
Every model built a game tracker app, then got the same follow-up change, word for word, in Taskade Genesis. Here is how each one did, best first.
In our tests, Claude Sonnet 5, DeepSeek V4 Pro, DeepSeek V4 Pro 0813, DeepSeek V4 Pro, high thinking, DeepSeek V4 Pro, max thinking, GPT-5.6 Luna, GPT-5.6 Sol, GPT-5.6 Terra, Grok 4.5, Grok 4.7, and Grok Build 0.1 all built a working app, and the follow-up change landed without breaking it. Each result comes from one to five builds, so read it as a direction, not a final rank.
Last test · 21 builds
The change
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.
Made to: Game tracker app. One Dota 2 player logs every match with the hero, the result and the stats. A dashboard shows win rate and streaks.
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.
The follow-up: Add a heroes page.
What does “works” mean here?
- The model first builds the app, then gets the follow-up change in the same chat.
- The change lands, and it stays in scope: nothing else is removed or redesigned.
- Everything that worked before still works: saved entries, required answers and earlier pages.
Each result covers the first build and the follow-up change together. Where the test recorded it, the finding says what happened to the change.
How did each model rank?
Works first, then works with gaps, then no app. Each result is one group, with its models listed by name.
| Model | |||
|---|---|---|---|
| Works · 11 | |||
Claude Sonnet 5 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. Works2 builds (same result) · Tested | Works | 2same result | |
DeepSeek V4 Pro A win-rate trend, a logged match saved, and the heroes page landed. About 15 sample matches showed as real. Works1 build · Tested | Works | 1 | |
DeepSeek V4 Pro 0813 Both settings saved a logged match and ranked heroes by games. One labeled its sample matches, and one showed them as real. Works2 builds (same result) · Tested | Works | 2same result | |
DeepSeek V4 Pro, high thinking Labeled its demo matches and kept them out of the stats. A logged match saved, and the heroes page landed. Works1 build · Tested | Works | 1 | |
DeepSeek V4 Pro, max thinking The best-looking tracker of its set, and a logged match saved. The follow-up message finished the interface and the heroes page. Works1 build · Tested | Works | 1 | |
GPT-5.6 Luna 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. Works2 builds (same result) · Tested | Works | 2same result | |
GPT-5.6 Sol Labeled its sample matches as a demo, saved a logged match with every detail, and put the new heroes page in both menus. Works1 build · Tested | Works | 1 | |
GPT-5.6 Terra 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. Works2 builds (same result) · Tested | Works | 2same result | |
Grok 4.5 Four pages with labeled sample matches, a logged match saved, and the heroes page landed. Its hero chart showed no data. Works1 build · Tested | Works | 1 | |
Grok 4.7 A three-page esports dashboard with labeled sample matches, a coach and a weekly automation. A logged match saved, and the heroes page landed. Works1 build · Tested | Works | 1 | |
Grok Build 0.1 Four pages with a win-rate trend, a logged match saved, and the heroes page landed. Its sample matches carried no label. Works1 build · Tested | Works | 1 | |
| Works, with gaps · 5 | |||
Claude Haiku 4.5 Saved logged matches and added the heroes page, but its top-heroes chart showed no data. Works, with gaps1 build · Tested | Works, with gaps | 1 | |
Claude Sonnet 4.5 Rejected every logged match and showed every sample match as a loss, so the win rate read 0%. Works, with gaps1 build · Tested | Works, with gaps | 1 | |
DeepSeek V4.1 Flash Five pages with labeled sample matches, but every hero showed as unknown and no match could be saved. Works, with gaps1 build · Tested | Works, with gaps | 1 | |
Grok 4.3 Saved a logged match and added the heroes page, but its average KDA card showed a garbled number. Works, with gaps1 build · Tested | Works, with gaps | 1 | |
Grok 4.20 Neither setting could log a match: one rejected every save, and the other built no app page while its summary described one. Works, with gaps2 builds (best has gaps, worst made no app) · Tested | Works, with gaps | 2best has gaps, worst made no app | |
Each row is one model version: its newest test on this request, the number of builds behind it, and what we found. Rows with the same result form one group and are listed by model name. A direction, not a final rank.
The first build of this app
More Code arena tests
No side-by-side results published yet: Change text and headings.
The Code arena →The full TSK-Bench ranking →Pick a model in TaskadePlans and pricingDownload the data (JSON)
How we grade these apps: read the full method.
FAQ
Which AI model adds a page to an app without breaking it?
In our tests, Claude Sonnet 5, DeepSeek V4 Pro, DeepSeek V4 Pro 0813, DeepSeek V4 Pro, high thinking, DeepSeek V4 Pro, max thinking, GPT-5.6 Luna, GPT-5.6 Sol, GPT-5.6 Terra, Grok 4.5, Grok 4.7, and Grok Build 0.1 all built a working app, and the follow-up change landed without breaking it. Each result comes from one to five builds, so read it as a direction, not a final rank.
How many AI models were tested on this change?
Models tested: 14, from 4 model families. Each got the same request, word for word. Some ran at more than one setting, for 16 results in all. 11 worked and 5 worked with gaps.
Can I add a page to an app in Taskade?
Yes. Ask for the change in the same Taskade Genesis chat as the first build, in your own words. Pick a model or let TSK-1 Auto choose.
