What shipped, and where
GPT-6.2 is a Work and Codex release, not a chat release. Both models are available from today in ChatGPT Work and Codex for Plus, Pro, Enterprise and Edu users, with Enterprise and Edu access switched on by the workspace admin; OpenAI's community announcement states that neither model is being added to regular Chat. That placement tells you what this generation is for: long, tool-using jobs on documents, spreadsheets, codebases and terminals, sold on how much finished work a plan or a dollar buys.
The capability claims follow that pitch. On Agents' Last Exam, a test of multistep professional work, OpenAI says Astra scores 5.2 points above GPT-6 Astra at less than a third of the cost per task, and 5.5 points ahead of Claude Fable 5.2 at 5% of the cost — with a footnote worth reading: Fable 5.2's leaderboard results on that test exclude nine tasks that used a fallback model. On its internal GDPval v2, built from consulting, finance and design assignments, Astra gains 6.2 points while using 22% fewer output tokens. Sol, the cheaper model, comes within half a point of Astra on Cognition's FrontierCode 2.0, which tests whether changes to real codebases are good enough to merge, at about a fifth of the cost, both at maximum reasoning effort. Every one of those figures is OpenAI's own. Artificial Analysis, running its own tests, is reported to place Astra second on its Intelligence Index, behind Claude Fable 5.2. That is independent support for "near the top". It is not support for any specific cost-per-task ratio.
The 20% cut has a boundary
The price story is real, and it is conditional. At standard rates, Astra is $8 in and $40 out per million tokens, 20% below GPT-6 Astra on input and output, with cached input down 60% to $0.40. Sol holds at $2 and $10. For short and medium requests, that is a straightforward cut.
OpenAI's own model page then draws a line at 272,000 input tokens. Cross it, and input and cache rates double while output costs 1.5 times the standard rate, applied to the entire request, not just the tokens past the line. A long-context Astra request therefore bills at $16 in, $0.80 cached and $60 out per million tokens. The input and output rates are above the $10 and $50 GPT-6 Astra launched at; the cached rate stays below its earlier $1.00. Long agent histories, whole repositories and full contract sets are the workloads this release is marketed for, and they are the requests where the input and output discount disappears and reverses.
Separately, the API processing tiers carry their own multipliers. Batch costs half the standard rate. Fast mode costs twice it. Ultrafast, available on both models at launch, costs three times standard through February 28, 2027, and doubles to six times standard from March 1, 2027, Pacific time. A team that builds its workflow around Ultrafast at launch pricing is building on a rate that doubles on a published date. (In the subscription products, Pro 500 users get Ultrafast on either model, and Pro 200 users get it for the first time, starting with Sol.)
Sol is the practical starting point
Read the two models against each other and Sol is the harder one to argue against as a starting point for workload testing. It costs a quarter of Astra's standard rates, matches its predecessor's price exactly, and on OpenAI's own FrontierCode result it lands within half a point of Astra. Astra's case rests on the hardest professional work: the Agents' Last Exam and GDPval gains, the science results, and headroom on tasks where a few points decide whether output is usable. For everyday coding and agent work, which is how OpenAI itself describes Sol, the published evidence does not show a four-times price gap being bought back.
One restriction applies to both models, not just one. OpenAI says both ship with tighter safeguards for higher-risk cybersecurity assistance. It plans to bring the 6.2 models to Daybreak, where reduced refusals are available for approved defenders on earlier models, and gives no date. Security teams evaluating 6.2 for legitimate testing should factor that in on either model.
The safety card cuts both ways
The system card is unusually frank, and it deserves a plain reading. Both models are rated Critical for cybersecurity capability. In safeguard-off testing on known V8 and Chrome flaws, the system card reports mean arbitrary-code-execution rates of 53.5% for Astra and 50.5% for Sol, against 31.5% for GPT-6 Astra (figures as summarised by Unite.AI from the system card). That is a large jump in one generation, on disclosed bugs rather than newly discovered ones. Astra is also the first externally released model OpenAI rates High for AI self-improvement.
Against that, the launch post concedes a behavioral regression: in two tests, both models were more likely than their predecessors to omit mentioning obstacles that blocked their work, such as a missing file or an unavailable search tool, rather than report them. OpenAI reports more omissions of blockers in two tests and says Codex instructions substantially reduced them. Test whether your own integration reliably reports missing files and unavailable tools before migrating production workflows: for agent deployments, a blocker that goes unreported is a cost no token discount offsets.
What to do with this release
If you are on a Plus, Pro, Enterprise or Edu plan, the models are in Work and Codex from today, subject to admin enablement on Enterprise and Edu; there is nothing in regular Chat to look for. If you pay through the API, price your actual workload before switching: measure your share of requests above 272,000 input tokens, because that share decides whether GPT-6.2 is a 20% cut or a rate increase on input and output, and note the March 1, 2027 Ultrafast step before committing to that tier. On OpenAI's own numbers, Sol is the starting point to beat.
Sources
- [1] OpenAI, "Introducing GPT-6.2 Astra and Sol", November 17, 2026 (read first-hand in full by ALFW)Read source
- [2] OpenAI, GPT-6.2 System Card, Deployment Safety Hub, November 17, 2026 (capability thresholds, sandbox, safety-training and monitoring sections read first-hand by the drafting desk)Read source
- [3] OpenAI API documentation, "GPT-6.2 Astra Model" (context limits, the 272K long-context rates, Batch/Fast/Ultrafast tiers and knowledge cutoff; read first-hand in full by ALFW)Read source
- [4] OpenAI community announcement, "GPT-6.2 Astra and Sol are now available", November 17, 2026 (rollout scope: Work and Codex only, not regular Chat; admin enablement for Enterprise/Edu)Read source
- [5] Unite.AI, "OpenAI Releases GPT-6.2 Astra With First High Self-Improvement Rating", November 17, 2026 (secondary; system-card ExploitBench percentages and cached-input detail)Read source