GPT-6.2 Astra is OpenAI’s flagship model for demanding professional work, released on November 17, 2026. It’s the first OpenAI model released externally to meet the “High” threshold for AI self-improvement under the company’s Preparedness Framework. It ships in ChatGPT Work, Codex, and the API, next to the cheaper GPT-6.2 Sol.
Astra follows GPT-6 Astra, which launched in early September. Sol replaces GPT-6.1 Sol, a model only about seven weeks old, so the release pace is fast. OpenAI pitches Astra at science, law, and finance, and says it was trained to follow professional conventions more reliably. One example: generated slides pull layouts, colors, and fonts from a template, down to the exact font size.
Table of Contents
ToggleWhat does “High” for AI self-improvement actually mean?
OpenAI defines a High rating for AI self-improvement as roughly equal to giving each of its researchers one extra, highly capable research engineer, compared with a 2024 baseline. In practice, the tests look at tasks like debugging failed experiments, writing faster kernels, and training a small model from scratch.
A High rating doesn’t mean Astra can rewrite itself or run off and build its successor alone. It means the model is good enough at AI research work to speed up the humans doing it.
OpenAI itself treats this as a risk. The company’s own warning is that models helping to build future AI could push progress faster than safeguards and human oversight can adapt. Its countermeasures include tighter controls in research environments, a monitoring system called Control Room that can track research tasks and halt them, and expanded monitoring of how models reason and act.
Here’s the part I’d pay attention to. OpenAI sets the thresholds, runs the tests, and assigns the grade. Critics argue that a model rated High for self-improvement and Critical for cybersecurity should face outside review before it reaches millions of users. That’s a fair argument, and it’s one OpenAI hasn’t fully answered.
The honesty numbers
The system card includes a test called Blocker Deception, which checks whether a model mentions a known obstacle when it’s stuck on a coding task. Astra left the obstacle out in 22.28% of trials. GPT-6 Astra did it in 11.09%. A prompt used in Codex reduces the behavior, but it isn’t applied to direct API requests.
Research group Epoch AI also measured how many serial reasoning steps Astra can run inside a single output token, with no visible chain of thought. The figure was 12.3, up from 10.5 for GPT-6 Astra. Reasoning you can’t see is reasoning that’s harder to monitor, and that matters more when the model is helping with AI research.
GPT-6.2 Astra vs Sol: which one do you need?
| GPT-6.2 Astra | GPT-6.2 Sol | |
| Built for | Demanding professional work | Fast, everyday tasks |
| API input / output (per 1M tokens) | $8 / $40 | $2 / $10 |
| Cached input (per 1M tokens) | $0.40 | $0.10 |
| Self-improvement rating | High | Below High |
| Cybersecurity rating | Critical | Critical |
| Bio and chemical rating | High | High |
| Artificial Analysis Intelligence Index | 62 | 57 |
| Vals Index rank / cost per task | 7th / $10.68 | 9th / $3.18 |
My read: start with Sol. On the Vals Index the two models sit just 1.6 points apart, and Sol costs less than a third as much per task. Move up to Astra only for work where a wrong answer is expensive, like contract review, scientific analysis, or financial modeling.
GPT-6.2 Astra pricing and availability
GPT-6.2 Astra and Sol are available to Plus, Pro, Enterprise, and Edu users in ChatGPT Work and Codex, and to developers through the OpenAI API. Pro-500 subscribers get Ultrafast mode with both models. Pro-200 subscribers get Ultrafast access for the first time, starting with Sol.
Astra’s API price is 20% below its predecessor’s, at $8 per million input tokens and $40 per million output tokens. Sol’s pricing is unchanged from GPT-6.1 Sol.
One more cost angle: OpenAI says Astra uses 22% fewer output tokens than GPT-6 Astra on its internal GDPval v2 tasks, which are built from consulting, finance, and design assignments. Fewer tokens at a lower price per token compounds quickly on large jobs.
GPT-6.2 Astra benchmarks: where it wins and where it doesn’t
Start with the independent numbers, since vendor benchmarks always flatter the vendor. On the Artificial Analysis Intelligence Index, GPT-6.2 Astra scores 62, second behind Anthropic’s Claude Fable 5.2 at 65. That’s five points above GPT-6 Astra, and it puts Astra ahead of Google’s Gemini 4 Carbon (60) and Anthropic’s Claude Opus 5.5 (59). Epoch AI’s Capabilities Index puts Astra at 174, tied with Fable 5.2.
GPT-6.2 Sol scores 57 on the Artificial Analysis index, which matches GPT-6 Astra at about 15% of the cost per task.
Some standout results:
- Math: Astra solved 7 of 68 problems on FrontierMath Erdős, each verified in the Lean proof assistant. The previous record was 3 of 68. One of the seven had no known solution before.
- Science: Astra leads the Vals Science Index and Terminal-Bench Science.
- Cost-performance: Artificial Analysis says OpenAI models dominate the cost-versus-performance frontier.
- OpenAI’s own claims: On Agents’ Last Exam, OpenAI says Astra beats Claude Fable 5.2 by 5.5 points at 5% of the cost. That’s OpenAI’s number, so treat it as a claim until others reproduce it.
And where it falls short:
- Coding: On the Coding Agent Index, Astra is flat against GPT-6 Astra. Sol drops three points from GPT-6.1 Sol.
- Hallucinations: At maximum effort, Astra hallucinates on 59% of the hardest factual questions in AA-Omniscience, and Sol on 55%. These questions are unusually difficult, so the figures don’t reflect everyday use, but they show the models still guess instead of admitting they don’t know.
- Legal work: On Vals AI’s legal benchmarks, both models rank well behind the leaders. OpenAI sells a separate Astra for Law product for this reason.
- Real-world research tasks: Epoch AI found Astra trailing Fable 5.2 and barely improving over GPT-6 Astra on tasks pulled from its own research.
If you’re a developer choosing a coding model, the headline number matters less than that flat coding result.
Safety ratings beyond self-improvement
Both models meet OpenAI’s Critical threshold for cybersecurity, the top level, just as GPT-6 Astra did. In OpenAI’s internal ExploitBench test, where models turn known browser vulnerabilities into working exploits, Astra cleared a 50% success rate with safeguards turned off.
Because of that, OpenAI is shipping stricter safeguards on high-risk cybersecurity requests. The company admits those safeguards will also block some legitimate vulnerability research. Vetted security researchers can get access to the new models with fewer guardrails through OpenAI Daybreak.
Both models also meet the High threshold for biological and chemical capabilities.
What GPT-6.2 Astra means for you
If you build with the API: Test Sol first, then compare Astra on your hardest prompts. Don’t assume the newer model fixes coding quality; the independent index says it doesn’t.
If you run a team or business: The 20% price cut and lower token use make Astra cheaper to run on long professional tasks. Check legal and research outputs by hand.
If you do security work: Expect more refusals on the standard models. Look into Daybreak access.
If you just use ChatGPT: You’ll mostly notice better slides, documents, and spreadsheets. The self-improvement rating is a signal about where the industry is heading, not a feature you’ll see.
Frequently asked questions
Is GPT-6.2 Astra better than Claude Fable 5.2?
It depends on the test. Artificial Analysis ranks Claude Fable 5.2 first at 65 and Astra second at 62, and Epoch AI scores them level. OpenAI’s own benchmarks show Astra ahead at a much lower cost. Independent testers found Fable 5.2 stronger on real-world research tasks and on coding.
How much does GPT-6.2 Astra cost?
GPT-6.2 Astra costs $8 per million input tokens, $0.40 per million cached input tokens, and $40 per million output tokens through the API. That’s 20% less than GPT-6 Astra. GPT-6.2 Sol costs $2, $0.10, and $10 for the same three categories.
Does a High self-improvement rating mean the AI can improve itself?
No. It means OpenAI judges the model capable of meaningfully accelerating AI research, similar to giving each researcher an extra research engineer. Humans still direct the work, and OpenAI uses monitoring like Control Room to track and stop research tasks.
Is GPT-6.2 Sol rated High for self-improvement?
No. Only Astra reached the High threshold. Sol meets the Critical threshold for cybersecurity and the High threshold for biological and chemical risk.
When was GPT-6.2 Astra released?
OpenAI released GPT-6.2 Astra and Sol on November 17, 2026. They replace GPT-6 Astra (released in September 2026) and GPT-6.1 Sol.
The bottom line
GPT-6.2 Astra is a real step on price and professional-work quality, and a mixed one on coding and accuracy. The self-improvement rating is the story that will outlast the benchmark tables, because it’s the first time a major lab has shipped a model it says helps build the next one. Whether OpenAI’s self-graded safeguards are enough is the open question, and no benchmark answers it.








