AI Model Reviews

Testing GPT-6.2 Astra for Web Design, 3D, UI and UX

In Codex Community's four tests, GPT-6.2 Astra built a tidy landing page and a useful dashboard. Its least convincing result took the longest: after roughly 45 minutes, a brief for a 3D space-agency website produced rendered images rather than a scene you could explore. That's a useful distinction if you're deciding which model to use for client work. A site can look finished in a screenshot and still miss the most important requirement.

Watch the Four-Test Review

Source: Codex Community's November 18, 2026 review and the supplied transcript. The times and design judgments are the presenter's, not results independently reproduced by JQ AI SYSTEMS.

What Astra Delivered in Four Tests

BriefReported timeWhat came back
Astra landing page18m 39sGood typography and spacing, but only about three sections. The presenter had seen much fuller treatments from Opus and Fable on the same brief.
TypeUI redesign25m 56sA clean page for TypeUI that looked too much like the previous build. The reviewer liked the type and colors but wanted a composition that felt specific to the product.
3D space-agency siteAbout 45mSatellite imagery gave the page a sense of depth, but the reviewer found rendered images in place of an interactive 3D experience.
Daily dashboardAbout 21mA single page for checking important information at a glance. The reviewer found it similar to Fable's result and slightly preferred Astra's visual treatment.

The first two builds show why polish and design range need separate checks. The launch page looked good but lacked depth. The redesign repeated a visual pattern that had worked on a different brief. More sections would not automatically make either site better. A stronger test is whether visitors can tell what the product does, find the details they need, and take the intended action.

The presenter selected high reasoning in ChatGPT Work for the landing page. Treat the times as observations from those particular runs, not as a prediction for another brief, model setting or tool.

The dashboard was a better fit for a familiar layout. Its job was to make information easy to scan, so originality mattered less. The presenter used Zapier MCP and described allowing actions such as drafting an email without granting permission to send it. For a pilot, check that each dashboard item traces back to the right source and use read-only access where possible. The video doesn't independently verify every item, tool call or connector.

Why the Benchmark and the Demo Can Both Be Right

OpenAI positions Astra for demanding projects and Sol for everyday work, with gains in output efficiency across both models. The Agents' Last Exam v1.1 leaderboard shows Astra at 78.3% on the scoring rubric and 47.5% on strict task passes in Codex. Fable 5.2's listed Claude Code run scored 72.8% with a 46.9% strict pass rate, and the leaderboard flags the exclusion of nine tasks that used a fallback model for that result. A rubric score and a strict pass rate answer different questions. These are also different model-and-tool setups, not a controlled head-to-head inside the same app.

The review is evidence about four creative and workflow tasks, not a test of unattended work over several hours. It also isn't a separate hands-on test of Sol. The useful conclusion is narrower: Astra was worth trying on the reviewer's bounded UI work, but efficiency didn't help when the requested artifact never arrived.

Give Your Next Design Test a Pass Condition

Pick one authorized project and compare Astra, Sol or Fable using the same briefs, assets and tool access where possible. Give every run the same time or spend cap. Judge visual direction separately from functionality: a 3D view must respond to input, the primary user flow must work on mobile, and dashboard data must match its sources. Log retries and human cleanup in the five-task scorecard before choosing the cheapest model that reliably passes.

JQ AI SYSTEMS starter prompt

Build and verify a design prototype

Ready to copy

Original editorial prompt, not the presenter's exact prompt. Launch claims and benchmark results checked November 22, 2026.

Video Chapters

TimeTopic
00:00GPT-6.2 Astra and the early reactions
00:30OpenAI's launch post
01:29Community examples on X
02:11Astra 6.2 benchmarks
03:52Test 1: Astra landing page
04:59Test 2: TypeUI redesign
05:57Test 3: 3D space-agency site
06:48Test 4: Daily dashboard with Zapier MCP
08:27Final verdict
Share
X LinkedIn Reddit
Build Yours

Want a system
like this one?

Book a free 30-minute call. We map your situation, identify the highest-impact automation, and figure out if we are a fit.

Book Free 30-min Call