GPT-6.2 Astra and Sol: OpenAI Brings Them to ChatGPT and Codex
OpenAI brings GPT-6.2 Astra and Sol to ChatGPT, Codex and the API: who can use them, token pricing, Ultrafast mode and what the company claims.

On November 17, 2026, OpenAI launched GPT-6.2 Astra and GPT-6.2 Sol, two AI models available in ChatGPT Work, in Codex and in the API for developers.12 Astra is designed for the most difficult projects, and Sol for everyday work. OpenAI claims that both complete tasks with fewer output tokens, so they allow users to do more before reaching the usage limit of their plan.13
Who can use the new models, and since when?
Both models are available immediately in ChatGPT Work and in Codex for Plus, Pro, Enterprise and Edu subscribers. In the API they are called gpt-6.2-astra and gpt-6.2-sol.14 The GPT-6 family began in September with Astra. Sol moved to version 6.1 a few weeks later, at DevDay.1 With this update, the $200-a-month Pro plan gets Ultrafast mode for the first time, but only with Sol. The $500 Pro plan has it on both models, as do eligible Enterprise and Edu workspaces, after an administrator enables it.15 According to OpenAI, Ultrafast generates tokens up to 4.5 times faster than the standard pace, but it consumes more of the usage limit. Speed is changed from the model selector.1
What are tokens and usage limits?
Tokens are the units of text that the model processes and on which pricing is based. Every ChatGPT plan has a usage limit, and a model that produces less output text to complete the same task theoretically leaves more room before reaching it.13 However, the lower number of output tokens is a claim made by OpenAI. Among the data cited, the company reports for Astra a gain of 6.2 points on GDPval v2, an internal test, with 22% fewer output tokens.2
How much do they cost in the API?
Astra costs $8 per million input tokens and $40 per million output tokens. These prices are 20% lower than those of GPT-6 Astra, which were $10 and $50.54 For text already sent and reused, the so-called cache, Astra costs $0.40 per million tokens, compared with $1 for the previous model.43 Sol keeps the prices of GPT-6.1 Sol: $2 per million input tokens, $0.10 for cache and $10 for output.53
What does OpenAI say about performance?
On Agents’ Last Exam, a multi-step professional work test, OpenAI claims 5.2 points more for Astra than GPT-6 Astra at less than a third of the cost. Compared with Anthropic’s Claude Fable 5.2, it claims a 5.5-point lead at about 5% of the cost.62 On Terminal-Bench-Science 0.1, which measures scientific work from the terminal, Astra gains 9 points over GPT-6 Astra. OpenAI calls it the best model in the world for scientists.6 Astra has also been trained to produce slides, spreadsheets and documents that better follow the conventions of professions such as law, finance and accounting. It can reproduce the design of a slide, from a file or a screenshot, including layout, colors and font size.16 In Codex, according to OpenAI, Astra works better at modifying and testing existing projects. For Sol, the company cites Cognition’s FrontierCode 2.0, which assesses whether changes to real code are good enough to be accepted by maintainers: at maximum reasoning effort, Sol remains half a point behind Astra at about a fifth of the cost.1 Basis, an AI startup for accounting and one of the first testers, reported that the average time for its long-running accounting jobs fell by 50%, and by up to 73% for the most complex ones.6
What do independent measurements show, and what risks are declared?
On the Artificial Analysis Intelligence Index, an independent index, Astra scores 62 points and is second behind Claude Fable 5.2, which has 65. The same index shows a hallucination rate of 59% for Astra and 55% for Sol on AA-Omniscience, and a 3-point drop for Sol on the Coding Agent Index.52 The safety documentation published by OpenAI shows that both models cross the Critical bar for cybersecurity and the High bar for biological and chemical capabilities, as defined in its Preparedness Framework. Consequently, the company has applied stricter default controls to the most dangerous prompts.3 Astra is also the first model released so far to hit the High level for AI self-improvement on OpenAI’s scale, so the company classes it as a risk. The competitor comparisons the company cites are drawn from OpenAI’s internal tests.53
This article was written by artificial intelligence from the sources cited, without human review. Automatic newsroom by Redazia.
Sources
Footnotes
-
OpenAI upgrades ChatGPT and Codex with new GPT-6.2 Astra and Sol models, 9to5Mac, 2026-11-17 ↩ ↩2 ↩3 ↩4 ↩5 ↩6 ↩7 ↩8 ↩9
-
OpenAI Releases GPT-6.2 Astra and Sol With a 20% Cheaper Astra API, MarkTechPost, 2026-11-17 ↩ ↩2 ↩3 ↩4
-
OpenAI lanzó GPT-6.2 y limitó su ayuda en ciberseguridad por el riesgo de que sirva para atacar, Infobae, 2026-11-18 ↩ ↩2 ↩3 ↩4 ↩5 ↩6
-
GPT‑6.2 Astra and GPT‑6.2 Sol, Simon Willison’s Weblog, 2026-11-17 ↩ ↩2 ↩3
-
OpenAI launches GPT-6.2 Astra and Sol with lower API prices and some benchmark gains, The Decoder, 2026-11-17 ↩ ↩2 ↩3 ↩4 ↩5
-
OpenAI rolls out GPT-6.2 Astra for professional work and GPT-6.2 Sol for everyday work, SiliconANGLE, 2026-11-17 ↩ ↩2 ↩3 ↩4


