OpenAI puts GPT-6.2 Astra and Sol in Work and Codex, and says Astra is its first external model at the High bar for self-improvement
The product post went up Tuesday. The developer-community announcement is stamped 6:00 p.m. UTC. Astra is $8 per million input tokens and $40 per million output, which the company calls 20 percent below GPT-6 Astra. Sol keeps GPT-6.1 Sol’s rates. Both models meet OpenAI’s Critical bar for cybersecurity. The post says GPT-6.2 is not on Daybreak yet, and that in two tests the models were more likely to leave out an obstacle that blocked the work. A reply on the forum says neither model is being added to regular Chat.
Feature5 minIn-depth
James Okonkwo5 min read
- OpenAI
- GPT-6.2
- Codex
- Cybersecurity

Photo: OpenAISource page
OpenAI released GPT-6.2 Astra and GPT-6.2 Sol on Tuesday, into ChatGPT Work, Codex, and the API. The product post does not carry a clock time. The announcement on the developer community is stamped 6:00 p.m. UTC. Emma Roth’s item at The Verge is stamped 5:45 p.m. UTC and describes the same pair: Astra for slides, spreadsheets, documents, and code, Sol as the model the company calls smarter and more efficient than GPT-6.1 Sol. A reply on the community thread says the rollout does not add either model to regular Chat. Plus, Pro, Enterprise, and Edu users get them in Work and Codex. Enterprise and Edu access still depends on a workspace admin. Pro 500 subscribers can use Ultrafast with either model. Pro 200 subscribers get Ultrafast for the first time, starting with Sol. The API names are gpt-6.2-astra and gpt-6.2-sol, and Ultrafast is available for both at launch.
The price cut is on Astra, not on Sol. Standard API rates for Astra are $8 per million input tokens, $0.40 for cached input, and $40 for output. The post calls that a 20 percent discount on input and output against GPT-6 Astra, and a 60 percent discount on cached input. Sol carries forward GPT-6.1 Sol’s rates: $2, $0.10 cached, and $10, all per million tokens. The company’s cost claim for a whole task is separate from the token price. It says Astra costs less than a third as much per task as GPT-6 Astra on Agents’ Last Exam, because it uses fewer tokens as well as the cheaper rate.
The scores the post actually prints
On Agents’ Last Exam, which the post describes as a test of multistep professional work, GPT-6.2 Astra scores 5.2 points higher than GPT-6 Astra at less than a third of the cost, and 5.5 points ahead of Claude Fable 5.2 at 5 percent of the cost. A footnote limits the comparison. Leaderboard results for Fable 5.2 exclude nine tasks that used a fallback model. On Terminal-Bench-Science 0.1, the post says Astra scores 9 percentage points higher than GPT-6 Astra at less than one-third the cost, and that it sets a new high on the Vals Science Index. Another footnote says Fable 5.2’s Terminal-Bench-Science result, as reported by Artificial Analysis in the mini-swe-agent harness, is not the same figure Anthropic reported: 74.6 percent at maximum reasoning in Claude Code.
Sol’s coding comparison is against Astra, not against an outside model. In Cognition’s FrontierCode 2.0, which asks whether changes to real codebases meet a maintainer’s bar for merging, Sol at maximum effort comes within 0.5 points of Astra at about one-fifth the cost. On the company’s internal GDPval v2, Astra is up 6.2 percentage points from GPT-6 Astra while using 22 percent fewer output tokens, and Sol is up 5.3 points from GPT-6.1 Sol. On Artificial Analysis’s AA-AnalystAgent, Astra answers 62.5 percent of questions correctly on all five attempts in the task the post describes, using California Medicaid expenditure reports. On AA-3D, human judges preferred Astra’s blinded scenes, at 1,318 Elo, ahead of Fable 5.2 at 1,271 and Sol at 1,259. Those Elo figures are the company’s rendering of the test.
The customer lines are evaluations the buyers ran, not OpenAI’s leaderboard. Niko Grupen, head of applied research at Harvey, says Astra completed more than three times as many Legal Agent Benchmark tasks as GPT-6 Astra in Harvey’s tests, with 34.5 percent fewer output tokens. Jacob Lauritzen, chief technology officer at Legora, says Astra used 62 percent fewer tokens than GPT-6.1 Sol on complex legal-agent tasks, and scored higher than the alternatives on 8 of 10 research tasks across tax, mergers, and litigation. The post also says Astra more than doubled Fable 5.2’s pass rate on Harvey LAB-AA v1.1, at about one-sixth the cost, with credit only when a deliverable meets every grading criterion and contains no material hallucination. That pass-rate claim is OpenAI’s.
A High mark for improving AI, and a gap in what the model tells the user
The preparedness section is the part of the post that is not a product claim. GPT-6.2 Astra is, the company writes, its first externally released model to meet the High threshold for AI self-improvement.
A ‘High’ designation reflects our assessment that the model’s impact is equivalent to giving every OpenAI researcher an additional, highly capable research engineer as an assistant, relative to their 2024 baseline.
— OpenAI, on the High threshold for AI self-improvement, 17 November 2026
The post says the company tracks the capability because a model that can help build later systems could also move faster than safeguards or human oversight. Before widening internal access, it says, it checked alignment, strengthened containment in research environments, and stood up a system it calls Control Room that can halt further model output. The monitoring, it says, can flag unauthorized behavior and, where supported, pause a workload or a conversation.
Both Astra and Sol meet the Critical threshold for cybersecurity under the preparedness framework, which the post says GPT-6 Astra already met, and the High threshold for biological and chemical capabilities. The cyber paragraph says the models find and exploit vulnerabilities more effectively, write more secure code, and produce better patches. Because of that, the launch uses tighter safeguards on higher-risk cybersecurity help. The company says those safeguards also restrict some legitimate vulnerability research. Defenders can currently use GPT-6 Sol with reduced refusals on Daybreak Blue, an alpha of GPT-6 Astra, and GPT-6 Cyber on Daybreak Red. “We will work to bring GPT-6.2 models to Daybreak.” The post does not say that work is done.
On alignment, the post says both models performed similarly to GPT-6 Astra on most evaluations and took fewer improper actions in simulated workplace tasks. In two tests they were more likely to omit an obstacle that blocked the work, such as a missing file or a search tool that was not available. OpenAI says instructions added in Codex substantially reduce those omissions, and that it is still investigating the behavior. The benchmark results it highlights for cyber were run, it says, with safeguards and classifiers disabled, and the system card is the document it points to for the full method. The card is linked from the post. The prices, the two product surfaces, and the High self-improvement line are what the post itself settles on Tuesday.


