I Gave GPT-5.6 Three Real Builds. Here Is Where It Won and Where It Broke.

When GPT-5.6 launched, I skipped the usual list of scores and gave it work.

I asked it to handle three very different builds:

1. Recreate the AI Builders landing page from screenshots

2. Build a native iPhone alarm app that makes you walk before it stops

3. Turn a voice-first chess idea into a playable 3D web game

Then I used ChatGPT Work to research a topic and turn the source material into social carousels.

This was not a clean laboratory test. It was closer to a normal workday: incomplete ideas, screenshots, bugs, taste decisions, and me changing direction while the agents worked.

That made it useful.

I gave GPT-5.6 screenshots of my AI Builders site and asked it to rebuild the page.

The first result was shockingly close.

The first landing-page result from the screenshot-driven build. Captured at 3:12.

The spacing, color system, hierarchy, calls to action, and overall feel were already there. I would still review responsiveness, accessibility, forms, analytics, disclosures, and every link before publishing. As a design-to-code first pass, it saved a huge amount of setup time.

The lesson was simple: give the model a clear target. A screenshot plus a real outcome beats a vague request for a "cool website."

The second build was Steppy, a native iOS alarm concept. When the alarm starts, the user has to walk a set number of steps before it can be dismissed.

Visually, the app looked good. The first physical-device test exposed the problem: the motion and fitness permission was not working correctly, so the step counter could not do its job.

I showed GPT-5.6 exactly what failed and asked it to repair the permission flow.

This is how I judge coding agents now. The first render is fun. The real test is whether the agent can understand a failure, inspect the implementation, and keep moving without destroying the rest of the app.

Steppy passed the visual test. It needed another engineering loop before it could pass the product test.

This one started as an idea I had in the car.

I wanted a beautiful 3D chessboard in the browser. You should be able to say a move, see a preview, and confirm or cancel it before the piece moves.

GPT-5.6 produced the strongest 3D chess interface I had seen from one of my prompts.

Voice Gambit running locally with the 3D board and voice-control panel. Captured at 12:25.

Then I tried it:

It was playable, impressive, and unfinished. All three can be true.

I also used ChatGPT Work to research posts on X, identify patterns, and create an AI Builders carousel.

One of the carousel assets produced from the research workflow. Captured at 14:08.

This is where GPT-5.6 felt bigger than a coding release. OpenAI describes the model as available across ChatGPT, Codex, and the API, and ChatGPT Work is designed to operate across apps and files to create finished materials.

I can start with a goal, research the inputs, build the software, inspect it in a browser or on a device, and turn the work into content without rebuilding the context in five separate tools.

| Build | First result | What worked | What still needed work |

| --- | --- | --- | --- |

| AI Builders site | Very strong | Visual hierarchy, layout, branding, one-shot fidelity | Production QA, forms, analytics, accessibility, compliance |

| Steppy iOS app | Strong visual prototype | Native UI, clear concept, fast first build | Motion and fitness permission, real-device validation |

| Voice Gambit | Best surprise | 3D design, voice commands, preview and confirmation | Chess logic and edge cases |

| Content workflow | Useful end to end | Research, synthesis, branded carousel output | Editorial judgment, sourcing, final voice review |

GPT-5.6 made all four workflows faster. It did not remove the operator.

The advantage goes to the person who can define the outcome, supply good references, test the result, explain a failure, and keep the agent focused.

For a quick demo, one prompt can be enough. For a business, I still want a proper brief, acceptance criteria, source files, QA, and approval rules.

The model is more capable. The quality of the final product still reflects the quality of the operating system around it.

Disclosure: Some links below are affiliate links. If you choose to buy through one of them, I may earn a commission at no extra cost to you.

Want to learn and build alongside other AI builders? Join us here.

This article documents my own build session. The projects shown were demonstrations and prototypes, not proof of production readiness, sales, revenue, or guaranteed results. Model access, product features, plans, limits, and pricing can change.