TSK-Bench · Code arena
Which AI model adds a field to a form without breaking it?
23 AI models built an intake form, then got the same follow-up change, word for word, in Taskade Genesis. Here is how each one did, best first.
Last test · 23 models · 34 builds
The change
Add one more question to a form the model already built. The new answer must save, and every earlier answer must still save.
Made to: Intake form. 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.
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.
The follow-up: Add one more question to the form.
What “works” means 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.
Ranked results
Works first, then works with gaps, then no app. Each result is one group, with its models listed by name.
| Model | |||
|---|---|---|---|
| Works · 13 | |||
Claude Opus 5 Saved every answer and both photos, and its email automation ran. The follow-up change landed. Works1 build · Tested | Works | 1 | |
Claude Opus 5.5 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. Works1 build · Tested | Works | 1 | |
Claude Sonnet 5 Both builds saved every answer and both photos, and the new teacher question landed. Sample families showed on the teacher page with no label. Works2 builds · Tested | Works | 2 | |
DeepSeek V4 Pro Saved every answer and both photos, and the new teacher question landed. Two sample families showed as real. Works1 build · Tested | Works | 1 | |
DeepSeek V4 Pro 0813 DeepSeek · Not in Taskade Both settings saved every answer and both photos, and the new teacher question landed. With max thinking, the follow-up turn finished the build. Works2 builds · Tested | Works | 2 | |
DeepSeek V4 Pro, max thinking 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. Works1 build · Tested | Works | 1 | |
DeepSeek V4.1 Flash Saved every answer and both photos, and its new-family automation ran. Three sample families showed as real. Works1 build · Tested | Works | 1 | |
GLM-5 Zhipu · Not in Taskade Saved every answer and both photos, and the new teacher question showed on its gallery page, but no link led to that page. Works1 build · Tested | Works | 1 | |
GPT-5.6 Sol 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. Works1 build · Tested | Works | 1 | |
GPT-6.1 Sol All three saved every answer and both photos, and every follow-up change landed. One teacher page misread its saved answers. Works3 builds · Tested | Works | 3 | |
Grok 4.5 Saved every answer and both photos, labeled its sample families, and added a class assistant and a weekly automation. The new teacher question landed. Works1 build · Tested | Works | 1 | |
Grok 4.6 Saved every answer and both photos, and the new teacher question landed. Its teacher page showed three sample families with no sample label. Works1 build · Tested | Works | 1 | |
Grok 4.7 Saved every answer and both photos, and the new teacher question landed. The 15-minute limit ended the build before its closing summary. Works1 build · Tested | Works | 1 | |
| Works, with gaps · 11 | |||
Claude Haiku 4.5 Saved every answer, but the family photo replaced the child photo on every submission. Works, with gaps1 build · Tested | Works, with gaps | 1 | |
Claude Sonnet 4.5 Saved every answer and both photos in a plain look, and its open teacher page listed parent phone numbers. Works, with gaps1 build · Tested | Works, with gaps | 1 | |
DeepSeek V3.2 DeepSeek · Not in Taskade Neither setting saved a family: one save was rejected, and the other sent the form to an address that does not exist. Works, with gaps2 builds · Tested | Works, with gaps | 2 | |
DeepSeek V4 Pro, high thinking The form dropped every photo a parent picked, so the two-photo rule blocked every submission. Works, with gaps1 build · Tested | Works, with gaps | 1 | |
GLM-4.7 Zhipu · Not in Taskade Saved every answer and both photos, but the teacher page crashed on load. Works, with gaps1 build · Tested | Works, with gaps | 1 | |
GPT-5.6 Luna One build saved every answer and both photos in under 5 minutes. In the other, every parent submission failed on the photo permission question. Works, with gaps2 builds · Tested | Works, with gaps | 2 | |
GPT-5.6 Terra 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. Works, with gaps3 builds · Tested | Works, with gaps | 3 | |
Grok 4.3 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. Works, with gaps1 build · Tested | Works, with gaps | 1 | |
Grok 4.20 xAI · Not in Taskade Neither build kept a parent's answers: one lost the child's name, and the other rejected every submission with no message. Works, with gaps2 builds · Tested | Works, with gaps | 2 | |
Grok Build 0.1 The photos landed, but 4 of 14 answers did not save, among them a required one, while the form thanked the parent. Works, with gaps1 build · Tested | Works, with gaps | 1 | |
Mistral Large 4 Mistral · Not in Taskade A polished form, but it dropped every photo a parent picked, so the two-photo rule blocked every submission. Works, with gaps1 build · Tested | Works, with gaps | 1 | |
| No app · 1 | |||
| No app | 1 | ||
Earlier tests (3)
The same request, word for word, in an earlier test of the same versions.
| Model | |||
|---|---|---|---|
Claude Sonnet 5 Saved every answer and both photos. The follow-up change landed. Works1 build · Tested | Works | 1 | |
GPT-5.6 Luna The most economical complete build: every answer and both photos saved. Its public assistant told a visitor a child's name. Works, with gaps1 build · Tested | Works, with gaps | 1 | |
DeepSeek V4.1 Flash 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. Works, with gaps1 build · Tested | Works, with gaps | 1 |
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 same-request table yet: Change text and headings.
The Code arena →The full TSK-Bench ranking →Download the data (JSON)
How we grade these apps: read the full method.
FAQ
Which AI model adds a field to a form without breaking it?
In our tests, Claude Opus 5, Claude Opus 5.5, Claude Sonnet 5, DeepSeek V4 Pro, DeepSeek V4 Pro 0813, DeepSeek V4 Pro, max thinking, DeepSeek V4.1 Flash, GLM-5, GPT-5.6 Sol, GPT-6.1 Sol, Grok 4.5, Grok 4.6, and Grok 4.7 all built a working app, and the follow-up change landed without breaking it. Each result comes from one to three builds, so read it as a direction, not a final rank.
How many AI models were tested on this change?
23 AI models from 7 model families got the same request, word for word. Some ran at more than one setting, for 25 results in all. 13 worked, 11 worked with gaps, and 1 did not produce an app.
Can I add a field to a form 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.