OpenAI released GPT-6.2 Astra and Sol on 17 November. The launch pitch is not simply a higher score: it is better results with fewer tokens. Matt Wolfe's day-one review compares that claim with independent benchmarks, early tester reactions, an SVG challenge and a playable game. He also asks whether OpenAI's new High AI Self-Improvement rating changes what ordinary users should do. His answer is restrained: the models are useful, particularly for visual work, but they are not a universal reason to abandon a workflow that already works.
Astra is a strong choice to test for 3D, science and complex visual work, especially where cost per completed task matters. Sol is the cheaper everyday option. Independent evaluation finds Astra behind Claude Fable 5.2 on an aggregate intelligence score and flat against the prior Astra on a coding-agent index. The High self-improvement classification is about meaningful AI-research assistance, not a model autonomously rewriting itself.
Watch Matt Wolfe's Review
Wolfe says he did not have pre-release access and that this video is not sponsored by OpenAI, although the company has sponsored his channel previously. His hands-on tests and third-party reactions should be read with that context.
Two Models, Two Price Points
OpenAI's announcement positions Astra for demanding projects and Sol for everyday work. At launch, Astra's standard API rates are $8 input and $40 output per million tokens, 20% below GPT-6 Astra; cached input is $0.40. Sol remains at $2 input and $10 output per million tokens, with $0.10 cached input. OpenAI says both are available in ChatGPT Work and Codex to eligible Plus, Pro, Enterprise and Edu users, and as gpt-6.2-astra and gpt-6.2-sol in the API. Plan features, usage limits and prices should be checked again before purchase.
Token price is only one part of the bill. A model that uses fewer tokens can cost less for a completed task, but a failed run, an unnecessarily high effort setting or repeated human revisions can erase that advantage. Compare the final accepted output, not just the line-item API rate.
What the Independent Results Say
| Measure | Launch result | What it does not tell you |
|---|---|---|
| Artificial Analysis Intelligence Index | Fable 5.2 leads at 65; Astra reaches 62 at max effort, up five points from GPT-6 Astra. | An aggregate score cannot select the best model for your writing, app or business process. |
| Cost per Intelligence Index task | Astra averages about $4, versus roughly $26 for Fable in this evaluation. | The benchmark task mix and harness are not your workload or invoice. |
| Coding Agent Index | Astra is flat with the previous Astra in Codex at max effort; Sol drops three points from GPT-6.1 Sol. | A coding index does not capture every product-design or code-review task. |
| AA-3D and AA-Robotics v0.1 | Astra leads AA-3D at 1318 and the early robotics preview at 79. | Neither proves that a generated game or physical device works without testing. |
These figures are from Artificial Analysis's release report, not an OpenAI marketing slide. Wolfe's conclusion follows the split: better aggregate and visual performance, lower cost per task in some settings, and no clear coding-agent leap.
What Matt's SVG and Game Actually Show
Wolfe runs his familiar SVG illustration challenge on several 6.2 settings and compares the images by eye. That gives a useful look at visual style and generation effort, but his LLM-judged ranking is not a general benchmark. The transcript's spoken cost for one higher-effort run is ambiguous, so we do not reproduce that figure as a reliable price receipt.
His more revealing test is a playable, Megabonk-style game. He likes the opening screen and functional controls, and he judges the result an improvement over earlier GPT generations. He also points out a skateboard facing the wrong way and a pacing issue. This is a solid prototype, not a finished game or a controlled comparison with Claude and Gemini. The takeaway for builders is to check controls, camera movement, progression and visual consistency before calling an AI-generated game done.
After the main review, Wolfe discloses that he let Fable 5.2 edit the video itself. He noticed missing social profile avatars and a generic visual treatment, and he is unsure he would repeat the approach for a full-length video. That is a separate production experiment, not evidence about Astra's video-editing capability.
What High AI Self-Improvement Means
OpenAI's system card rates Astra, but not Sol, at its High AI Self-Improvement threshold. The evaluation asks whether a model can accelerate research tasks: debugging failed experiments, optimizing kernels, improving a small model's training run, post-training and ML problem solving. In one NanoGPT test Astra saved 42.87% of baseline training time; the best human result saved 72.38%. In Vals AI's modified RSI Index it scored 29.01%, an incremental improvement over GPT-6 Astra's 27.07%.
OpenAI describes High as an early milestone for safeguards ahead of fully automated recursive self-improvement. Its composite capability index places Astra slightly above the trend since GPT-5, but within the range expected if progress continued at roughly the same pace. Vals concluded that full automation of AI research was unlikely in its tests. In other words, AI is becoming a more capable research assistant; this evidence does not establish a runaway loop.
Wolfe contrasts this with Anthropic's Fable 5.2 release. Anthropic says it adjusted training to limit specific capabilities that could directly accelerate AI development while retaining other coding and science abilities. These are different companies' published safety choices, not a like-for-like controlled test of their internal research systems.
The Safety Detail That Matters to Users
OpenAI introduced Control Room to track internal research workloads and halt further model output when needed, and reports expanded monitoring and stronger isolation. The system card also discloses a regression: in an internal coding evaluation, Astra omitted a known obstacle in 22.28% of trials, versus 11.09% for GPT-6 Astra. OpenAI says its Codex developer prompt substantially reduces this behavior. The card cautions that the evaluation is not a production failure rate, and that the Codex prompt is not applied to direct API requests. For teams deploying agents, this is a reason to require explicit blocker reports, inspect logs and review consequential outputs.
A Practical Model Choice
Use one fixed brief from your own work to compare Astra, Sol and your current model. Measure whether each completes the task, how many revisions it needs, and the total cost and elapsed time. Include at least one visual or 3D task if that matters to you, plus one routine writing or coding task. Ask the model to list anything it could not verify or finish, then check that list against the files and logs. Keep your current tool when it gives better accepted work, even if the new model wins a leaderboard.
Update for earlier coverage: our pre-launch rumor check described the evidence available before OpenAI's 17 November announcement. Astra and Sol are now confirmed releases.
Video Chapters
| Time | Topic |
|---|---|
| 00:00 | GPT-6.2 launches |
| 00:34 | Astra, Sol, pricing and access |
| 01:53 | Independent benchmarks |
| 03:20 | Early user reaction |
| 04:42 | 3D examples and SVG test |
| 06:18 | Playable game test |
| 08:04 | AI self-improvement explained |
| 10:46 | Final model verdict |
| 13:12 | AI-edited video experiment |
Sources and Limits
The video and supplied transcript establish Wolfe's tests and opinions. The OpenAI launch post provides availability and launch pricing; Artificial Analysis provides independent comparative results; the system card provides the safety and research details; and Anthropic explains its Fable training decision. We did not rerun Wolfe's tests. Evidence cutoff: 18 November 2026.