What OpenAI shipped
OpenAI released two models on 17 November 2026, US time. GPT-6.2 Astra is the flagship, aimed at the hardest professional work. GPT-6.2 Sol is the cheaper, faster model for everyday coding and writing. The OpenAI developer forum announcement gives the API IDs as gpt-6.2-astra and gpt-6.2-sol. It also says the models appear in ChatGPT Work and Codex for Plus, Pro, Enterprise and Edu plans, and that the rollout doesn't add either model to regular Chat. Codex CLI users have to update to the latest version before the new models show up.
The release lands about a week after Anthropic's Claude Fable 5.2, and 9to5Google frames it as OpenAI's reply. For anyone building on the API, the useful details are in OpenAI's model pages, and the caveats come from the independent evaluators.
Specs and pricing
Both models share the same shape on paper. Each has a 1,050,000-token context window. Input is capped at 922,000 tokens and output at 128,000 tokens. Both work with the Chat Completions, Responses and Batch endpoints, and neither supports Realtime, Assistants or fine-tuning. The knowledge cutoffs differ: Astra's is 18 May 2026 and Sol's is 1 June 2026. Both models always reason before answering. Reasoning tokens are the hidden working the model produces before its visible answer, and you're billed for them as output. The API rejects requests that set reasoning effort to none, and neither model page lists the other effort levels.
Standard prices per million tokens, in US dollars:
- Astra: $8 input, $0.40 cached input, $40 output, $10 for a 30-minute cache write
- Sol: $2 input, $0.10 cached input, $10 output, $2.50 for a 30-minute cache write
The Decoder reports that Astra's price is 20% below GPT-6 Astra's, and that Sol keeps GPT-6.1 Sol's price. Batch jobs cost half the standard rate, which suits overnight evaluation runs or bulk document processing.
The headline prices leave out two costs. Both model pages say that once a request goes over 272,000 input tokens, input and cache rates double and output costs 1.5 times the normal rate for the whole request. A long-context Astra call therefore costs $16 per million input tokens and $60 per million output tokens. Anyone who plans to fill the 1M window with a whole codebase or a stack of contracts should budget at those rates. Ultrafast mode is the other cost. 9to5Mac reports OpenAI's claim that Ultrafast generates tokens up to 4.5 times faster than standard. In the API, Ultrafast costs three times the standard rate until 28 February 2027 and six times from 1 March 2027 Pacific time, which is 9pm NZDT on 1 March here. Fast mode costs twice the standard rate.
Rate limits are the same for both models. The Build tier allows 5,000 requests and 1,000,000 tokens per minute, and the Grow tier allows 15,000 requests and 40 million tokens per minute.
Benchmarks, and who ran them
Most of the headline claims come from OpenAI's own evaluations. OpenAI says Astra beats GPT-6 Astra by 5.2 points on Agents' Last Exam at under a third of the cost per task, and that Sol comes within half a point of Astra on Cognition's FrontierCode 2.0 at about a fifth of the cost. FrontierCode 2.0 tests whether a model's changes to real codebases are good enough for maintainers to merge. OpenAI also cites an internal measure, GDPval-v2, where Astra gains 6.2 percentage points and uses 22% fewer output tokens. Sources for these figures are The Decoder's summary and the forum post.
The independent results, as The Decoder reports them, are more mixed. On the Artificial Analysis Intelligence Index, Astra scores 62, behind Fable 5.2 on 65, and Sol scores 57. Artificial Analysis also found that coding has stalled. Astra is level with GPT-6 Astra on the Coding Agent Index, and Sol dropped three points from GPT-6.1 Sol. On Artificial Analysis's deliberately hard factual question set, Astra hallucinated in 59% of answers at maximum effort and Sol in 55%.
Vals AI places Astra seventh and Sol ninth on its overall index, with Sol the cheapest model in the top 10 at $3.18 per task. Astra costs $10.68 per task, about half of GPT-6 Astra's $21.28, but it still averages 91,000 output tokens per task. Vals ranks Astra first on its science index. On legal work the models drop sharply: Astra is 30th and Sol 52nd on Vals' Legal Research benchmark. Epoch AI gives Astra 174 on its Capabilities Index, the same as Fable 5.2, but finds Astra trails Fable 5.2 on real-world research tasks.
Safety settings that change behaviour
OpenAI rates both models at its "Critical" cybersecurity threshold and has tightened refusals on higher-risk security help. The forum post says OpenAI is working to bring the models to Daybreak, its programme for approved defenders, with fewer refusals. Teams building penetration-testing or vulnerability-triage tools should test their prompts again before switching models. The Decoder also reports that in two of OpenAI's alignment tests, the models more often failed to report blockers such as missing files or unavailable tools. An agent that keeps going quietly when a tool fails is harder to debug, so explicit tool-error checks in your harness matter more with these models.
What we'd do with it
For most New Zealand teams paying in US dollars, Sol is the model to test first. It keeps last generation's price, OpenAI's own coding numbers put it close to Astra, and Vals measured it as the cheapest top-10 model per task. If your product already runs on GPT-6.1 Sol, swapping the model ID and re-running your own evaluation set is a small job. Watch the Coding Agent Index drop, and compare agentic coding results on your own repositories before you commit.
Astra makes sense for science-heavy or document-heavy work where Vals and OpenAI agree it leads. It's a weaker choice for legal research, where its independent rankings are poor despite OpenAI's Harvey claims. Keep prompts under 272,000 input tokens wherever retrieval can do the job, because crossing that line doubles the input bill. Cache stable system prompts at a tenth of the input price. If you've built a latency-sensitive feature on Ultrafast, budget for the price doubling in March, since the six-times rate will make Ultrafast Astra output cost $240 per million tokens.
Some details remain unknown. OpenAI's announcement page didn't load for us, the model pages don't list the reasoning effort levels, and the independent benchmark results reached us through The Decoder's report. If you benchmark either model on a local workload, such as te reo Māori text or NZ legislation, we'd like to see the results on our projects page.
What we checked, and assumptions
We picked this story because neither ASAI Insights nor DataForge Insights had covered GPT-6.2 Astra and Sol in their feeds as of 18 November 2026. OpenAI's announcement page at openai.com returned an error when we tried it, so the specs and prices come from OpenAI's developer model pages and the official forum post. The independent benchmark figures (Artificial Analysis, Vals AI, Epoch AI) come from The Decoder's report, and we couldn't open the evaluators' own pages for these models. The 9pm NZDT time is our conversion of midnight Pacific Standard Time.
Sources
- GPT-6.2 Astra model page, OpenAI API docs
- GPT-6.2 Sol model page, OpenAI API docs
- GPT-6.2 Astra and Sol are now available, OpenAI Developer Community
- OpenAI launches GPT-6.2 Astra and Sol with lower API prices and some benchmark gains, The Decoder
- OpenAI upgrades ChatGPT and Codex with new GPT-6.2 Astra and Sol models, 9to5Mac
- OpenAI releases GPT-6.2 Astra to regain performance lead from Claude Fable 5.2, 9to5Google
- ASAI Insights feed
- AI Model Release Tracker feed, Opper
- New AI Models timeline, LLM Gateway
Build with people working on the same things
Join the Auckland Society for AI: NZ$599 a year with a ForgeBox and the members' projects area. Our Auckland meetups are open to everyone.