GPT-6.2 Astra and Sol: API prices, context window and what independent tests show
OpenAI's new models cost $8/$40 and $2/$10 per million tokens; independent testers rank Astra second to Claude Fable 5.2 but find coding flat.
· 7 min read
Written by AI from the sources linked below. How we use AI
On this page
OpenAI released GPT-6.2 Astra and GPT-6.2 Sol on 17 November 2026, two models for the API, ChatGPT Work, Codex and GitHub Copilot. Astra is priced at $8 per million input tokens and $40 per million output tokens; Sol stays at $2 and $10. Independent testers rank Astra second on one broad index, behind Anthropic's Claude Fable 5.2, but their coding results are flat, and OpenAI says Astra is the first model it has released at its "High" level for AI self-improvement.
Key points
- GPT-6.2 Astra costs $8 per million input tokens, $0.40 cached and $40 per million output tokens; GPT-6.2 Sol costs $2, $0.10 and $10 (OpenAI API model pages).
- Both models take text and images, return text, and have a 1,050,000-token context window with up to 128,000 output tokens.
- Artificial Analysis scores Astra 62 on its Intelligence Index, behind Claude Fable 5.2 at 65, and Sol 57; Epoch AI puts Astra at 174 on its Capabilities Index, tied with Fable 5.2.
- Artificial Analysis finds no coding gain: Astra is level with GPT-6 Astra and Sol drops three points on its Coding Agent Index.
- OpenAI says both models reach its "Critical" cybersecurity threshold, and that Astra is the first externally released model at its "High" self-improvement threshold.
What happened
OpenAI's API documentation lists two new model IDs, gpt-6.2-astra and gpt-6.2-sol. It describes Astra as "our most capable model for your hardest projects" and Sol as the "go-to model for everyday work, from coding to powering agents at scale". Both accept text and image input and produce text only; audio and video are not supported, and fine-tuning is not supported. Both reject requests that set reasoning effort to none. Tools in the Responses API include web search, file search, code interpreter, hosted shell, apply patch, computer use, MCP and tool search, and a Multi-agent beta lets a model delegate work to subagents.
According to The Decoder's report on OpenAI's announcement, Astra and Sol are available to Plus, Pro, Enterprise and Edu users in ChatGPT Work and Codex, and Pro-500 subscribers can use an Ultrafast mode with both models, while Pro-200 subscribers get Ultrafast access starting with Sol. We could not open OpenAI's announcement page itself (it requires JavaScript), so claims below that come from the announcement are marked "OpenAI says" and are as reported by The Decoder.
GitHub made both models generally available in Copilot the same day, for Copilot Pro+, Max, Business and Enterprise users, with a gradual rollout across Visual Studio Code, Visual Studio, Copilot CLI, the Copilot coding agent, the Copilot app, github.com, GitHub Mobile, JetBrains IDEs, Xcode and Eclipse. GitHub says that in early testing both models caught significantly more real issues in code review than earlier GPT-6 models while using fewer tokens, and that Sol caught nearly twice as many as GPT-6.1 Sol. On Business and Enterprise plans, new models are enabled by default unless an administrator has turned off the global default or disabled them.
Background
Astra is the premium line in OpenAI's GPT-6 family; Sol is the mid-priced workhorse. This release is a point update: GPT-6.2 follows GPT-6 Astra and GPT-6.1 Sol. OpenAI's pricing is the main headline. The Decoder reports OpenAI's statement that Astra's API price is 20 percent below its predecessor's, while Sol keeps GPT-6.1 Sol's pricing. We have not checked the predecessor's price against OpenAI's page, so we quote only the new figures.
The release lands in a crowded week for AI model news. We covered how OpenAI's GPT-6 Cyber scored on ExploitGym earlier, and how ChatGPT Work and Codex now share one usage pool, the subscription context for this launch.
The numbers
Price and limits. From OpenAI's model pages, per million tokens: Astra is $8.00 input, $0.40 cached input and $40.00 output, with cache writes at $10.00 for 30 minutes of retention. Sol is $2.00 input, $0.10 cached, $10.00 output, and $2.50 for cache writes. For requests with more than 272,000 input tokens, input and cache rates double and output costs 1.5 times the short-context rate for the full request. Batch costs 50 percent of standard rates. Fast mode costs twice standard; Ultrafast costs three times standard through 28 February 2027 and six times from 1 March 2027. The knowledge cutoff is 18 May 2026 for Astra and 1 June 2026 for Sol. Artificial Analysis confirms the $8 and $40 Astra pricing and the 1M-token context window.
Intelligence rankings. Artificial Analysis' Intelligence Index rolls up nine evaluations. As reported by The Decoder, Claude Fable 5.2 leads at 65, GPT-6.2 Astra scores 62, Google's Gemini 4 Carbon 60, Claude Opus 5.5 59 and GPT-6.2 Sol 57.
Cost per task. Vals AI, as reported by The Decoder, measures $10.68 per task for Astra, $3.18 for Sol and $21.28 for GPT-6 Astra, and ranks Astra 7th and Sol 9th on the Vals Index.
| US$ per task | |
|---|---|
| GPT-6 Astra | 21.28 |
| GPT-6.2 Astra | 10.68 |
| GPT-6.2 Sol | 3.18 |
Source: Vals AI, via The Decoder
Other independent results. Epoch AI's homepage lists GPT-6.2 Astra as the top model on its Epoch Capabilities Index at 174; The Decoder says that ties Claude Fable 5.2. On FrontierMath Erdős, a set of unsolved problems checked in the Lean proof assistant, Astra solved 7 of 68, against a previous record of 3 of 68 held by GPT-6.1 Sol, and one of the seven had no known solution before (The Decoder, citing Epoch AI).
What it means
Developers. Sol is the interesting model for most teams: OpenAI says on FrontierCode 2.0 it trails Astra by just 0.5 points at one-fifth the cost, and Vals AI measured its per-task cost at under a third of Astra's. Astra's benefit is concentrated in science, professional documents and 3D or design work. OpenAI says it leads on Terminal-Bench-Science and, on the AA-3D head-to-head judged by humans, scores 1,318 Elo against 1,271 for Claude Fable 5.2 and 1,259 for Sol.
Coding teams should be careful. Artificial Analysis finds both models flat on its Coding Agent Index (measured with the Codex harness at maximum compute): Astra stays where GPT-6 Astra was, and Sol falls three points, from a 14-point drop on SWE-Atlas-QnA-AA where it omits key facts more often. Vals AI places Astra 5th and Sol 7th on Terminal-Bench 4. GitHub's claims about code review are early-testing results from GitHub, not independent benchmarks.
Accuracy. Artificial Analysis flags a high hallucination rate on its AA-Omniscience test: 59 percent for Astra and 55 percent for Sol at maximum effort. The test uses unusually hard questions, so The Decoder cautions that the numbers do not reflect everyday hallucination rates. Epoch AI found Astra weaker than Fable 5.2 on real-world research tasks and barely improved on GPT-6 Astra there. Vals AI found legal weak spots: Astra ranks 11th on HLAB and 30th on its Legal Research benchmark, with Sol 21st and 52nd.
Safety and security researchers. OpenAI says both models reach the "Critical" cybersecurity threshold in its Preparedness Framework, as GPT-6 Astra did, and adds stricter safeguards for risky security requests that it says will also limit some legitimate security testing. It says it will offer the models through OpenAI Daybreak to vetted researchers. Its internal ExploitBench figure, over 50 percent for Astra, comes from tests with safeguards turned off. OpenAI also says Astra is the first externally released model to hit its "High" threshold for AI self-improvement, comparable to giving each researcher an extra skilled research engineer. Epoch AI measured Astra at 12.3 serial reasoning steps within a single output token, up from 10.5 for GPT-6 Astra, which means less visible chain of thought for monitors to inspect.
What we don't know yet
- We have not read OpenAI's announcement page or the GPT-6.2 system card published on 17 November; all announcement claims here come via The Decoder.
- OpenAI's headline cost comparisons (for example, an advantage of "5 percent of the cost" on Agents' Last Exam) use OpenAI's own benchmarks and cost definitions.
- Real-world coding quality: independent coding indices are flat, and early reports are mixed.
- Whether the stricter cybersecurity safeguards will block ordinary developer work.
What to watch next
Watch for the system card's detail on the self-improvement and cybersecurity findings, rate-limit changes as demand arrives, and independent coding results on your own repositories. Teams on Copilot Business or Enterprise should check whether their administrator wants these models on by default.
Sources
- OpenAI API: GPT-6.2 Astra model page and GPT-6.2 Sol model page (primary)
- OpenAI: Introducing GPT-6.2 Astra and Sol (announcement; page could not be opened, claims as reported by The Decoder)
- The Decoder: OpenAI launches GPT-6.2 Astra and Sol with lower API prices and some benchmark gains
- GitHub Changelog: OpenAI's GPT-6.2 Astra and GPT-6.2 Sol now available
- Artificial Analysis: GPT-6.2 Astra
- Epoch AI benchmarks hub
Written by
The official account of the people who run audiotale.net: announcements, conversation starters and the newsroom's drafts. Posts written with AI say so. Every other account here is a real person.
Read next
GPT-6.1 Terra: US$1/US$5 per million tokens, on Bedrock and Copilot
OpenAI's new mid-tier model targets computer use and everyday agentic coding at half Sol's price; the benchmark numbers are OpenAI's own.
audiotale.net team6 min read
OpenAI models cut AMD GPU design PPA by up to 14%, Jalapeño area 13%
OpenAI says its models improved heavily optimised AMD GPU circuits, with some changes already in a design sent to manufacturing, and is building GPT-Synopsys with Synopsys.
audiotale.net team6 min read
OpenAI to embed six Apollo Research evaluators for a four-month safety pilot
Apollo Research will work inside OpenAI during training and internal deployment to test whether models try to subvert their own safeguards, with published reports and stated limits.
audiotale.net team7 min read