Language Models

Seven Tests Later, GPT-6.2 Astra Only Clearly Beat Claude Fable 5.2 at Building a Single 3D Character

November 19, 20260 views阅读中文原文
Seven Tests Later, GPT-6.2 Astra Only Clearly Beat Claude Fable 5.2 at Building a Single 3D Character

The short version

On the day GPT-6.2 Astra and GPT-6.2 Sol went live, the Chinese AI reviewer T8star-Aix ran a 25-minute hands-on comparison: seven tasks, every model pushed to the Extra High reasoning setting, with Claude Fable 5.2 as the reference point.

His verdict fits in three lines:

  • It can take over part of Claude's workload. The clearest case is a one-shot 3D mascot, where he says Astra beat Fable 5.2 outright.
  • Video and code-driven animation still lag. Astra closed the gap from "a world apart" to "roughly on par", but did not pass Claude there.
  • Price worked in OpenAI's favour. Astra costs less than the previous GPT-6 Astra, and Sol did not go up, so "the same work for fewer tokens" holds.

This is not a scored benchmark. Each of the seven tasks ran once, and the quality calls are the reviewer's own visual judgment. Treat it as field notes from someone who actually used the models, not a leaderboard.

What was tested: two model tiers, seven tasks

OpenAI released two models on 17 November: GPT-6.2 Astra for heavy project work and GPT-6.2 Sol for everyday tasks. Both carry a 1.05-million-token context window. On list pricing, Astra is about 20% cheaper than the previous flagship, while Sol sits level with GPT-6.1 Sol.

T8 deliberately skipped scoreboards and picked tasks that produce something you can look at:

  1. Code animation — one cycling-animation prompt, rendered as HTML in 24 different drawing styles, then exported as video.
  2. 3D scene — generate an HTML scene and let the model edit it, to see whether the taste holds up.
  3. Half-open brief — a low-altitude racing game starring a crane, with everything else left to the model.
  4. Keynote deck — an invented brand, and the model decides how to structure a full launch presentation.
  5. Browser music workstation — the model writes a page that can make music in the browser.
  6. Desktop pet / 3D character — the task Fable 5.2 won against both GPT models last round, retested here.
  7. The finale: robot football — three robots per side, each model writing one JS strategy, with inertia, a bouncing ball, dribbling by pushing the ball, 20 ms per step, memory carried across matches, a requirement to circle behind the ball before kicking, and only one robot allowed in the box. Twenty matches per pairing, 60 in total.

Task by task: one clear win, and it matters

Code animation (task 1). The reviewer calls the improvement obvious. At Extra High, output is close to Fable 5.2, and the 6.2 generation finally has real code-animation ability. He notes Sol's visual taste trails Astra but is still acceptable.

3D scene (task 2). The result splits cleanly: Sol still trails Fable 5.2 by a visible margin, while Astra is "genuinely close" on 3D — and on looks alone he prefers Astra.

Flying game (task 3). All three did well, with strong stylisation and most of the requested mechanics present. No separation here.

Keynote deck (task 4). All three were mediocre. His explanation: a slide deck simply has fewer effects to show off than code animation, so nothing stood out.

Browser music workstation (task 5). All three "looked roughly the part". His judgment is restrained — a professional tool generated in one pass is usable at best, and real work needs iteration.

Desktop pet / 3D character (task 6). This is the highlight of the video. Astra's one-shot output was "identical to the source", good enough to drop into a video as a finished model. He says it beat Fable 5.2 outright, even though Fable also ran at its highest reasoning setting this time.

Robot football (task 7). After reviewing 60 matches, Fable 5.2 is still the strongest, which he calls beyond doubt. But Astra improved massively over the previous round, where it was outclassed, and now counts as very good. With the price cut factored in, he considers the value proposition solid.

The most telling line in the video is about timing, not scores

T8 offers an observation that carries more signal than any of the test results: OpenAI sat on this release for a long time while X and group chats complained. His read is that Astra was still too far behind Fable 5.2 and Opus 5.5, and shipping early would have invited mockery — so OpenAI waited until the gap narrowed.

He also mentions something personal: his laptop would not boot before recording, Fable 5.2 could not fix it, and Astra eventually did. One anecdote proves nothing about capability, but it explains why he closes by saying GPT really is strong this time.

How to use it

  • General users get both tiers inside ChatGPT. Sol for everyday work, Astra for long documents and complex projects.
  • Developers and agent builders can reach them through Codex and the API under the model IDs gpt-6.2-astra and gpt-6.2-sol. Tool use — function calling, web search, computer use — requires the Responses API; Chat Completions does not support it.
  • Anyone wanting to reproduce the tests can grab the code and music, which the reviewer says are fully open-sourced in the description and comments on Bilibili and YouTube.

Where it fits

  • One-shot 3D characters and props. The only category where the video declares a clear win. Useful for short-video assets and placeholder models.
  • Code animation and visual pages. The gap is now small enough that anyone tired of Claude account instability can work with this.
  • Keynote decks and document deliverables. It will produce them, but do not expect anything striking; all three models merely passed.
  • Everyday coding and batch jobs. Sol is the value pick here — officially only a fraction behind Astra on coding benchmarks at a far lower cost.

Where it does not fit

  • Video pipelines that need reliable output. The reviewer's own experience is that video and code animation still trail, so test in small batches before switching.
  • Professional tools that must be right the first time. The browser music workstation needed iteration; a single pass produced a prototype.
  • Model selection by benchmark score. This video has no scoring system and runs each task once; it cannot stand in for a benchmark.
  • Unattended agents whose output you cannot verify. A model saying it finished is not proof that it finished.

Pitfalls when reading this kind of comparison

  1. Mistaking a one-shot success for reproducibility. Every result is a single generation, with no retry counts and no failure rates.
  2. Ignoring the reasoning setting. Everything here ran at Extra High; default settings will likely underperform the video.
  3. Treating taste as capability. "I prefer Astra's look" is a preference, not a measurable difference.
  4. Forgetting it is one person's view. The reviewer says plainly at the end that he hopes Chinese vendors catch up — the stance is declared, not hidden.

Cost and next steps

At list prices, Astra costs roughly five times Sol for the same token volume. So the practical move is not to switch everything to Astra:

  1. Pick the three tasks you run most, try both Sol and Astra on each, and compare pass rates rather than first impressions.
  2. Reserve Astra for the areas where it clearly leads, such as 3D characters and modelling assets.
  3. If your pipeline depends on video or code animation, stay on your current model and first rerun the same prompt with the open-sourced code from this video.
  4. Pin the reasoning setting to whatever you tested with, or the comparison means nothing.

Sources

  • Bilibili video (T8star-Aix, 17 November 2026): https://www.bilibili.com/video/BV1cfsm6vEVG
  • OpenAI announcement: https://openai.com/index/introducing-gpt-6-2-astra-sol/
  • The Decoder report: https://the-decoder.com/openai-launches-gpt-6-2-astra-and-sol-with-lower-api-prices-and-some-benchmark-gains/

The video content is based on a local speech-to-text transcript, since the video has no subtitle track. Official specifications and pricing come from OpenAI's announcement page and media reports. This article is compiled from multiple media reports.

Sources

Comments

0 published
Log inorsign up to leave a comment.

Loading comments...