GPT-6.2 Astra costs 20% less than GPT-6 Astra, matching Claude Fable 5.2. Sol stays at $2/$10.
OpenAI's GPT-6.2 Astra and Sol arrived November 17 in ChatGPT Work, Codex, the API and GitHub Copilot. Who gets them, what changed in pricing, OpenAI's benchmarks, and how Claude and Gemini list prices compare.
November 18, 2026

OpenAI released GPT-6.2 Astra and GPT-6.2 Sol on November 17, 2026, in ChatGPT Work, Codex and the API. Its launch post describes Astra as "A frontier model built for your hardest projects" and Sol as "An efficient and versatile model for everyday work." Astra costs $8 per million input tokens and $40 per million output tokens, 20% below GPT-6 Astra. Sol stays at $2/$10, the price of GPT-6.1 Sol.
What changed from GPT-6 Astra and GPT-6.1 Sol
Astra follows GPT-6 Astra, since OpenAI cancelled GPT-6.1 Astra in September, according to The Next Web. Sol follows GPT-6.1 Sol. From OpenAI's pricing page and model pages:
- Astra price: besides the 20% cut, cached input drops from $1.00 to $0.40 and 30-minute cache writes from $12.50 to $10. Requests over 272K input tokens cost $16/$60, down from $20/$75.
- Sol price: unchanged, with $0.10 cached input and the same long-context surcharge.
- Discount tiers: Batch is still half price, but there is no Flex rate. The pricing page says "We won't be adding Flex to new models." Ultrafast costs 3x Standard through February 28, 2027 and 6x from March 1; GPT-6 Astra and GPT-6.1 Sol Ultrafast cost 6x now.
- Limits: both keep a 1.05M-token window, 922K maximum input and 128K maximum output. Knowledge cutoffs move to May 18, 2026 (Astra) and June 1, 2026 (Sol).
- API changes: reasoning effort "none" is rejected and tool calling requires the Responses API. The API changelog adds that 6.2 changes how Responses keeps reasoning when earlier turns are edited and how it treats the top-level instructions field.
Who gets it in ChatGPT, Codex and the API
- ChatGPT Work: OpenAI's help page lists Plus, Pro, Business, Enterprise and Edu, on web, desktop and mobile. Enterprise and Edu admins must enable the models first. New Work chats default to Sol (Light effort on consumer plans, Medium in Enterprise); pick Astra under Advanced.
- Free and Go: not on OpenAI's list. Those ChatGPT plans run GPT-6 Luna.
- Codex: both are in the Codex app and CLI, and Codex CLI 0.171.0 makes Sol the default catalog choice. Codex and Work share one usage allowance.
- Ultrafast: Pro $500 gets it on both models, and Pro $200 gets it on Sol, a first for that tier. 9to5Mac reports output up to 4.5 times faster, using more of your allowance.
- API: gpt-6.2-astra and gpt-6.2-sol on Chat Completions, Responses and Batch.
OpenAI's benchmarks, and what outside testers found
These figures are vendor-reported. OpenAI's comparison table, as reproduced by AI Catchup, lists Astra, Sol and Claude Fable 5.2, in that order:
- Agents' Last Exam v1.1: 78.3%, 73.0%, 72.8%, at $2.89, $1.32 and $60.05 per task. OpenAI's Fable run excludes nine tasks that used a fallback model.
- Harvey LAB-AA v1.1 (legal): 15.0%, 11.1%, 7.2%.
- FrontierCode 2.0: 44.2%, 43.8%, 49.1%. Sol is within half a point of Astra at about a fifth of the cost, and Fable leads both.
- Artificial Analysis Intelligence Index v5.0.1: 62.1, 57.3, 64.6.
- Against older models: Astra gains 6.2 points on GDPval v2 over GPT-6 Astra with 22% fewer output tokens, per Digital Trends, and 9 points on Terminal-Bench-Science 0.1.
The Next Web reports that Epoch AI found Astra roughly tied with Fable 5.2 on its Capabilities Index, but averaging 65% on realistic work tasks, 17 points below Fable. It quotes developer Theo Browne, who tested Astra before launch: "I would not choose this over Opus 5.5 for my work as a developer."
The system card says both models meet OpenAI's Critical threshold for cybersecurity, so higher-risk security help is restricted by default. Both "were more likely to omit known obstacles when reporting their work," which OpenAI says the Codex developer prompt substantially reduces.
How the price compares with Claude and Gemini
List prices per million tokens, from our model database and, for Neon, Google's pricing page:
- Claude Fable 5.2: $8/$40, cache reads $0.20, one rate across 1M.
- Claude Opus 5.5: $4/$20, cache reads $0.20.
- Gemini 4 Neon: $3/$15, cached $0.15, one rate across 1M.
- Gemini 4 Carbon: $2/$10 through November 30, then $4/$20.
- Claude Sonnet 5.5: $2/$10, the same as Sol.
Astra matches Fable per token but charges twice as much for cache reads, and its rates rise above 272K input tokens. For a month with 400 million uncached input, 2 billion cached and 40 million output tokens, in requests under 272K and before cache writes:
- GPT-6 Astra: $4,000 + $2,000 + $2,000 = $8,000.
- GPT-6.2 Astra: $3,200 + $800 + $1,600 = $5,600, 30% less.
- Claude Fable 5.2: $3,200 + $400 + $1,600 = $5,200.
- Opus 5.5, or Carbon from December 1: $1,600 + $400 + $800 = $2,800.
- Gemini 4 Neon: $1,200 + $300 + $600 = $2,100.
- GPT-6.2 Sol: $800 + $200 + $400 = $1,400.
Token use differs by model, so per-task costs can reverse this order. See our Fable 5.2 post and the pricing index.
GitHub Copilot has it, Cursor does not
- GitHub Copilot: the GitHub changelog says both are available on Copilot Pro+, Max, Business and Enterprise, billed at provider list pricing, in the IDEs, Copilot CLI and the coding agent. Rollout is gradual. GitHub says Sol caught nearly twice as many real code review issues as GPT-6.1 Sol in early testing.
- Cursor: OpenAI said in August it would end its model supply to Cursor, proposing a November 12 shutoff, InfoWorld reported. Cursor's models page lists OpenAI models up to GPT-5.6 and no GPT-6 model. OpenAI's help center points Cursor users to the Codex IDE extension.
Before you switch, and what OpenAI has not said
- On GPT-6 Astra or 6.1 Sol: check for code that sets reasoning effort to "none" or calls tools through Chat Completions.
- On Flex: Batch is the only discount on the 6.2 models.
- On long prompts: price requests over 272K separately; Fable 5.2 and Neon keep one rate.
- On security work: expect more refusals. OpenAI says it is "working to bring these models to Daybreak so approved defenders can access them with reduced refusals."
The changelog lists no deprecation, and OpenAI has not given a retirement date for GPT-6 Astra or GPT-6.1 Sol, said whether Free and Go users will get the 6.2 models, or said when the models reach Daybreak.
Some links in this article are affiliate links. Learn more.