What does GPT-6.2 Astra buy you when the job is a hard visual build? Pat Simmons's no-hype review puts it against GPT-6 Astra, Claude Fable 5.2 and Claude Opus 5.5 on three briefs: a WebGL-heavy website, a browser baseball game and a walkable mid-century house. Fable takes all three quality wins in Simmons's rankings. The new Astra is the least expensive on every run, but its cheapest result is also the clearest miss.
GPT-6.2 Astra cost $57.82 in API-equivalent usage across the three tests and took 178 minutes of combined build time. It ranked second, fourth and third. Fable 5.2 ranked first each time and cost about $1,082. That makes Astra worth testing for a draft under a tight budget, not a substitute for checking that the result meets the brief.
Watch Pat Simmons's Full Review
What the Comparison Controls
Simmons's methodology used high effort in each vendor's own agent: Codex for the Astras and Claude Code for the Claude models. He used broadly comparable creative tooling and ranked anonymous builds before revealing the models. There was one run per model per task, and the older Astra and Opus baseball entries were reused from October with an earlier version of the brief. The rankings are one reviewer's preferences under those conditions, not an isolated test of model weights or an estimate of win rates.
Test 1: Zero University's WebGL Site
The reference was Why Zero, an interactive site built around scroll-driven scenes. The brief asks for original assets and fidelity in the ice, 3D hand, drawing interaction and transitions. Simmons used his replication playbook. At the rank reveal, he preferred Fable's hand, 3D elements and motion. Astra 6.2 came close enough for second at $26.95 and 51 minutes, against Fable's $392.74 and 265 minutes. Compare how the interactions behave, not just how closely the first frame matches.
Test 2: A Game Has to Play Like a Game
The baseball brief called for an original draft roster and a full game against the computer, including pitching, batting, fielding and base running. Fable's Porchlight League won the game comparison. Simmons judged Astra 6.2's run a miss on the requested gameplay. The older Astra placed above it and finished eight minutes sooner for $0.35 more. On this job, the lowest cost didn't buy the better result, and the video does not establish why the newer run fell short.
Test 3: The Walkable Mid-Century House
This brief asked for an original 1958 desert home with a walkable interior and exterior, changing light and a cinematic tour. At the house reveal, Simmons liked Astra 6.2's pool reflection and mountain backdrop but found the rooms and furniture less developed than the top entries. Fable won for lived-in period detail, with Opus second. Astra's version took 95 minutes at $25.41; Fable took 590 minutes at $660.43. The unanswered question is how those scenes would compare at the same time or cost limit.
Open All 12 Builds and Compare Their Receipts
Each table is in Simmons's ranked order. The links go to the public artifacts on AI for Mortals; an open demo is not proof its full brief passed. Input and output figures are the rounded token counts reported in the source receipts.
Zero University
| Rank / live build | Cost / time | Input / output tokens |
|---|---|---|
| 1. Fable 5.2 | $392.74 / 265 min | 610.1M / 3.3M |
| 2. GPT-6.2 Astra | $26.95 / 51 min | 41.1M / 172k |
| 3. Opus 5.5 | $180.63 / 388 min | 701.0M / 995k |
| 4. GPT-6 Astra | $152.44 / 207 min | 119.8M / 376k |
Backyard baseball
| Rank / live build | Cost / time | Input / output tokens |
|---|---|---|
| 1. Fable 5.2 | $28.62 / 91 min | 32.1M / 363k |
| 2. Opus 5.5 | $25.56 / 120 min | 57.0M / 405k |
| 3. GPT-6 Astra | $5.81 / 24 min | 2.4M / 52k |
| 4. GPT-6.2 Astra | $5.46 / 32 min | 5.6M / 56k |
Mid-century house
| Rank / live build | Cost / time | Input / output tokens |
|---|---|---|
| 1. Fable 5.2 | $660.43 / 590 min | 533.7M / 4.4M |
| 2. Opus 5.5 | $166.71 / 471 min | 409.1M / 1.9M |
| 3. GPT-6.2 Astra | $25.41 / 95 min | 37.6M / 144k |
| 4. GPT-6 Astra | $140.01 / 230 min | 88.7M / 395k |
Reported totals: GPT-6.2 Astra $57.82 / 178 min; GPT-6 Astra $298.26 / 460 min; Fable 5.2 about $1,082 / 946 min; Opus 5.5 $372.89 / 979 min. These are API-equivalent estimates in USD from runs made on subscriptions, not invoices or all-in project budgets. The totals are as published; displayed runs are rounded.
Pat's Exact Prompts and Shorter Test Briefs
Open the exact original under each project on AI for Mortals. The copy buttons here provide shorter JQ AI SYSTEMS adaptations, not the original benchmark text. Use the originals for a closer reproduction, and use only reference material you are authorized to use.
Buy a Finished Result, Not the Cheapest Attempt
OpenAI's launch rates for Astra 6.2 are $8 input, $0.40 cached input and $40 output per million tokens at Standard short-context pricing. Its Agents' Last Exam claims concern a different suite from these creative builds. For your own test, specify one time or spend cap per model and a minimum acceptance bar. A game must complete an inning; a walkthrough needs a route a new user can navigate. Repeat the close cases, then compare model cost and human repair time for the builds that actually pass.
Bonus: Pat's Separate Fable 5.2 Blender Test
Simmons points to this separate look at Fable in Blender at the end of the review. It is not a fourth scored round.
Video Chapters
| Time | Segment |
|---|---|
| 00:00 | Introduction |
| 00:48 | Test 1: Zero University website |
| 04:44 | Website build A |
| 11:17 | Website build B |
| 15:28 | Website build C |
| 21:18 | Website build D |
| 25:12 | Website rankings and costs |
| 26:45 | GPT-6.2 benchmarks and pricing |
| 33:00 | Test 2: Backyard baseball |
| 34:09 | Baseball build A |
| 36:27 | Baseball build B |
| 37:35 | Baseball build C |
| 38:16 | Baseball build D |
| 38:42 | Baseball rankings and costs |
| 39:58 | Test 3: Mid-century modern house |
| 40:57 | House build A |
| 43:13 | House build B |
| 45:48 | House build C |
| 47:46 | House build D |
| 49:25 | House rankings and costs |
| 51:14 | Final verdict |
Sources and Caveats
The embedded review and supplied transcript provide the rankings and hands-on comments. The 17 November AI for Mortals write-up supplies the original prompts, public artifacts and receipts. We checked the links and reconciled the reporting with the video; JQ AI SYSTEMS did not rerun or score the builds. Evidence cutoff: 18 November 2026.