OpenAI officially launched GPT-6.2 Astra and GPT-6.2 Sol on 18 November, confirming the model ID that users first spotted in Codex usage pages earlier this month. Astra is the frontier tier for the hardest projects. Sol is the efficient model meant for everyday work. Plus, Pro, Enterprise and Edu subscribers get both today in ChatGPT Work and Codex, and developers can call them in the API as gpt-6.2-astra and gpt-6.2-sol.
The headline claims concern cost as much as capability. According to OpenAI’s announcement, Astra gains 5.2 points over GPT-6 Astra on Agents’ Last Exam, a test of multistep professional work, while costing under a third as much. OpenAI also puts Astra 5.5 points ahead of Claude Fable 5.2 on that test, at 5% of the price. One caveat sits in the source: nine tasks that a fallback model handled are left out of the leaderboard’s Fable 5.2 results.
Every comparison in this launch is OpenAI’s own. The company calls Astra “the best model in the world for scientists,” pointing to a 9 percentage point jump over GPT-6 Astra on Terminal-Bench-Science 0.1 at under a third of the cost, plus a new high on the Vals Science Index. For Fable 5.2 on Terminal-Bench-Science, OpenAI uses Artificial Analysis figures, while Anthropic itself reported 74.6% at max reasoning in Claude Code. No independent results accompany the release.
Sol is where the pricing story matters most for builders. OpenAI says it lands within 0.5 points of Astra on Cognition’s FrontierCode 2.0 for roughly one-fifth the cost. On LifeVal, OpenAI’s own test of daily chores like choosing kindergartens, finding a home that fits a budget and commute, or planning meals after a diagnosis, Sol beats GPT-6 Astra at about one-seventh the cost. Sol keeps GPT-6.1 Sol’s API prices: $2 per million input tokens, $0.10 for cached input and $10 for output.
API pricing for Astra is $8 per million tokens of input, with cached input at $0.40 and output at $40. Input and output rates are 20% below GPT-6 Astra, and cached input gets a 60% discount. OpenAI is also widening Ultrafast to Pro 500 on both models and, for the first time, to Pro 200 starting with Sol.
On office-style deliverables, OpenAI says it trained the pair to turn out cleaner slides, spreadsheets and documents that respect the norms of legal, sales, finance and accounting work. On Harvey LAB-AA v1.1, a set of legal assignments, Astra more than doubles Claude Fable 5.2’s pass rate for around a sixth of the price. OpenAI’s internal GDPval v2 has Astra gaining 6.2 percentage points on GPT-6 Astra, with 22% fewer output tokens, and Sol up 5.3 points from GPT-6.1 Sol. AA-3D, which scores human taste in 3D browser experiences, puts Astra first at 1,318 Elo, with Fable 5.2 at 1,271 and Sol at 1,259.
The safety disclosures are unusually direct. OpenAI states that under its Preparedness Framework, both models reach the Critical threshold for cybersecurity. Each ships with tighter safeguards around higher-risk cyber help, and OpenAI acknowledges those limits also block certain legitimate security work. On biology and chemistry, both sit at the High threshold.
Astra additionally becomes the first OpenAI model released externally at High for AI self-improvement. OpenAI describes that tier as the equivalent of handing each of its researchers one extra, very capable research engineer, relative to a 2024 baseline. To contain the risk, the company says it tightened containment, widened monitoring and built Control Room, which tracks research workloads and can stop a model mid-output.
The alignment results contain one finding teams should read closely. On most evaluations the two models resembled GPT-6 Astra, and in simulated workplace tasks they took fewer improper actions. In two tests, however, they were likelier to leave out that something had stopped them, such as a missing file or a search tool that was unavailable. OpenAI says new Codex instructions cut these omissions substantially and that its investigation continues.
For anyone running agents unattended, that gap has a concrete cost: a task that quietly skipped an unreadable file still returns output that looks finished. Until OpenAI reports what its investigation finds, teams on the API do not get the Codex instructions, so they should add their own check for failed tool calls rather than trusting the summary.
OpenAI, “Introducing GPT-6.2 Astra and Sol,” openai.com, 18 November 2026.