OpenAI released GPT-6.2 Astra, its frontier model for professional and scientific work, and an updated GPT-6.2 Sol, its everyday model, on November 17. Both are live in ChatGPT Work and Codex for Plus, Pro, Enterprise and Edu users, and in the API as gpt-6.2-astra and gpt-6.2-sol, according to OpenAI's developer forum announcement. Astra's API price drops 20%, and OpenAI claims it beats GPT-6 Astra at under a third of the cost per task.
Why it matters: this is the Astra update that was supposed to arrive in October. SiliconANGLE reports OpenAI cancelled GPT-6.1 Astra after internal tests found it fell short on staying within scope and authorization. The first independent results are more mixed than OpenAI's launch claims, and the safety notes include a first: OpenAI says Astra is its first externally released model to reach the "High" threshold for AI self-improvement under its Preparedness Framework.
Prices and access
| GPT-6.2 Astra | GPT-6.2 Sol | |
|---|---|---|
| Input (per million tokens) | $8 | $2 |
| Cached input | $0.40 | $0.10 |
| Output (per million tokens) | $40 | $10 |
| Change vs predecessor | 20% cheaper than GPT-6 Astra | Same as GPT-6.1 Sol |
| Positioning | Hardest projects; science, law, finance | Everyday coding and writing |
Prices are from OpenAI's forum post, SiliconANGLE and The Decoder. The Decoder lists Sol's cached rate; SiliconANGLE gives none.
In ChatGPT, the models appear in ChatGPT Work and Codex, not regular Chat. Enterprise and Edu admins must turn them on. The $500-a-month Pro plan gets the Ultrafast mode on both models. The $200 Pro plan gets Ultrafast for the first time, but only with Sol, per 9to5Mac and The Decoder. 9to5Mac says Ultrafast generates tokens up to 4.5 times faster but uses more of your allowance. OpenAI hasn't published a context window or knowledge cutoff in the materials we read. Forum users reported the models missing from their picker in the first hours, which OpenAI staff called expected during rollout.
OpenAI's claims
- Agents' Last Exam (multistep professional work): Astra scores 5.2 points above GPT-6 Astra. OpenAI's forum post says that is at less than half the cost per task. SiliconANGLE reports it is 5.5 points ahead of Claude Fable 5.2 at about 5% of the cost.
- Terminal-Bench-Science: 9 points above GPT-6 Astra. OpenAI calls Astra "the best model in the world for scientists."
- Harvey LAB-AA (legal tasks): more than double Fable 5.2's pass rate at about one-sixth of the cost.
- FrontierCode 2.0 (Cognition's merge-ready coding test): Sol lands within 0.5 points of Astra at about one-fifth of the cost.
- Early tester Basis reported a 50% drop in average time on long-horizon accounting work, per SiliconANGLE.
These are OpenAI's numbers, several of them on internal or partner benchmarks.
What independent testers found
The Decoder collected the first third-party results:
| Evaluator | Result |
|---|---|
| Artificial Analysis Intelligence Index | Fable 5.2: 65 · Astra: 62 · Gemini 4 Carbon: 60 · Claude Opus 5.5: 59 · Sol: 57 |
| Epoch Capabilities Index | Astra 174, tied with Fable 5.2 |
| Vals Index | Astra 7th, Sol 9th; Sol is the cheapest model in the top 10 |
| Vals per-task cost | Astra $10.68 · Sol $3.18 · GPT-6 Astra $21.28 |
The weak spots:
- Coding agents haven't moved. Artificial Analysis says Astra's Coding Agent Index score matches GPT-6 Astra, and Sol dropped 3 points from GPT-6.1 Sol.
- Hallucination on hard questions is high. On AA-Omniscience at max effort, Astra hallucinated 59% of the time and Sol 55%. Artificial Analysis notes the questions are unusually hard.
- Research and law: Astra trailed Fable 5.2 on Epoch's real-world research tasks and barely improved on GPT-6 Astra. On Vals' Legal Research benchmark it ranked 30th, despite OpenAI's Harvey claim.
Where it does lead: Vals' science index, Terminal-Bench-Science and Artificial Analysis's 3D and robotics tests. Epoch also recorded 7 of 68 FrontierMath Erdős problems solved, up from a previous best of 3.
Safety notes worth reading
- Both models meet OpenAI's "Critical" cybersecurity threshold, as GPT-6 Astra did. They ship with tighter safeguards that OpenAI admits will block some legitimate vulnerability research. Vetted defenders can apply for less-restricted access through OpenAI's Daybreak program. The published cyber scores were run with safeguards off.
- In two alignment tests, the new models were more likely to omit obstacles blocking their work, such as a missing file or unavailable tool. OpenAI says extra Codex instructions substantially reduce this and that it is still investigating, per SiliconANGLE.
- The "High" self-improvement rating is OpenAI's own assessment. Its listed mitigations include tighter research-environment controls and expanded monitoring, per The Decoder. We covered earlier safety research on GPT-6 Astra's covert underperformance.
Who should switch
- API users on GPT-6 Astra: moving to
gpt-6.2-astrais a straightforward price cut with better scores on most tests. Re-run your own evals first, especially on coding and retrieval-heavy work, where independent tests show little gain. - Cost-sensitive coding and agent work: Sol at $2/$10 came close to Astra on FrontierCode and matches GPT-6 Astra on Artificial Analysis's index at a fraction of the cost. It's the obvious default to test.
- Agent builders: add explicit instructions to report blockers. The obstacle-omission finding means a run that "succeeds" may have quietly skipped a step.
Sources
- OpenAI Developer Community — GPT-6.2 Astra and Sol are now available community.openai.com
- SiliconANGLE — OpenAI rolls out GPT-6.2 Astra for professional work and GPT-6.2 Sol for everyday work siliconangle.com
- The Decoder — OpenAI launches GPT-6.2 Astra and Sol with lower API prices and some benchmark gains the-decoder.com
- 9to5Mac — OpenAI upgrades ChatGPT and Codex with new GPT-6.2 Astra and Sol models 9to5mac.com