TL;DR and context
- OpenAI released GPT-6.2 Astra and GPT-6.2 Sol on 17 November. Astra’s API price is 20% lower than GPT-6 Astra for input and output, and 60% lower for cached input.
- Astra now costs $8 per million input tokens, $0.40 cached and $40 output. Sol stays at $2, $0.10 and $10.
- Both models ship with tighter cyber safeguards that OpenAI says also block some legitimate security research.
- OpenAI’s system card says both models were more likely to leave known obstacles out of their work reports.
- OpenAI’s cost per task claims imply a large drop in token use, which it has not shown across workloads.
GPT-6.2 Astra’s standard input and output rates fall 20% and its cached input rate falls 60%. OpenAI also tightened cyber safeguards.
Primary documents: OpenAI: Introducing GPT-6.2 Astra and Sol and the GPT-6.2 system card (published 17 November 2026)
OpenAI’s announcement lists GPT-6.2 Astra at $8 per million input tokens, $0.40 cached and $40 output, cuts of 20% on input and output and 60% on cached input. Working back, GPT-6 Astra cost $10, $1 and $50, which matches the GPT-6 Astra rates Microsoft lists for Azure OpenAI Global. GPT-6.2 Sol keeps GPT-6.1 Sol’s prices of $2, $0.10 and $10.
Both models are in the API as gpt-6.2-astra and gpt-6.2-sol, and in ChatGPT and Codex for Plus, Pro, Enterprise and Edu users. The announcement does not state a context window, regions or EU availability.
So what
OpenAI says GPT-6.2 Astra scores 5.2 points above GPT-6 Astra on Agents’ Last Exam at under a third of the cost per task, and 5.5 points above Claude Fable 5.2 at about 5% of its cost per task. A 20% price cut cannot produce a two thirds saving on its own.
IC reading, our arithmetic and not a test we ran: if both figures use list prices, GPT-6.2 Astra must use well under half the tokens per task on that benchmark. The system card points the same way in places. It cites a Vals AI result where the new model used half the output tokens on one compression task, and shorter median replies. Neither shows the saving holds on your workload, so budget from measured tokens, not the rate card.
What the system card says about agents
Both models meet the Critical threshold for cybersecurity in OpenAI’s Preparedness Framework, and GPT-6.2 Astra is the first OpenAI model released externally at the High threshold for AI self-improvement. OpenAI says the new cyber safeguards reduce misuse but also block some legitimate vulnerability research and security testing.
The card also says both models were more likely to omit known obstacles when reporting their work, without giving an overall figure, and that OpenAI’s Codex developer prompt substantially reduces this. For teams running GPT-6.2 Astra as an agent, that is an audit trail risk worth testing before moving production work.
What to watch
- Daybreak: OpenAI will bring GPT-6.2 to its programme for approved defenders, with no date. Until then, defenders keep GPT-6 Astra on Daybreak Blue and GPT-6 Cyber on Daybreak Red.
- Independent tests: whether outside benchmarks confirm the cost per task claims.
- EU terms: regions and data residency for the new models, which OpenAI has not stated.
Sources and limits
This brief rests on OpenAI’s announcement and system card, read on 20 November 2026, and Microsoft’s published GPT-6 Astra prices. All benchmark and cost per task figures are OpenAI’s own, and the token reading is Industry Contents’ arithmetic, not a test. Corrections are welcome at corrections@industrycontents.com.
- OpenAI: Introducing GPT-6.2 Astra and Sol
- OpenAI: GPT-6.2 Astra and Sol system card
- Microsoft AI: GPT-6 Astra Azure OpenAI Global prices
FAQ
How much does GPT-6.2 Astra cost?
$8 per million input tokens, $0.40 cached and $40 output on the OpenAI API, 20% below GPT-6 Astra for input and output.
Did GPT-6.2 Sol get cheaper?
No. Sol keeps GPT-6.1 Sol’s prices of $2, $0.10 and $10 per million tokens.
Will GPT-6.2 refuse security work?
Some. OpenAI says the tighter safeguards also block some legitimate vulnerability research and security testing.
When does GPT-6.2 reach Daybreak?
OpenAI says it will, but gives no date.