
The Library Media — Science, Technology & Health · AI News & Intelligence Desk
Cheaper at the Top, Rated Higher for Self-Improvement: OpenAI Ships GPT-6.2 Astra and Sol
By Catalina Clarke
OpenAI released two new models on Tuesday, and the pair pull in opposite directions. GPT-6.2 Astra, the company’s flagship, costs a fifth less than the model it replaces. At the same time, OpenAI’s own system card for the release rates Astra at the High level for AI self-improvement — the first model OpenAI has deployed externally that reaches that threshold under its Preparedness Framework — and rates both new models Critical for cybersecurity, the framework’s most serious level.
GPT-6.2 Astra is built for the hardest professional tasks; GPT-6.2 Sol is the smaller, efficient model for everyday coding and writing. Both are available in ChatGPT to all paid plans, with limited access for free users. In Codex, Astra is now the default model for paid users and Sol is also available. Through the API, the models are named gpt-6.2-astra and gpt-6.2-sol.
The headline numbers are the prices, charged per million tokens. Astra costs $8 for input, $0.40 for cached input and $40 for output — 20% lower than GPT-6 Astra, by OpenAI’s account, and, as launch coverage noted, exactly the list price of Anthropic’s Claude Fable 5.2. Sol costs $2 for input, $0.10 for cached input and $10 for output, which OpenAI describes as 20% cheaper than GPT-6 Sol.
The fine print sits on OpenAI’s model pages. Both models take a context window of 1,050,000 tokens, with up to 922,000 tokens of input and 128,000 of output. And for very long requests — those over 272,000 input tokens — the input and cache rates double and output is billed at one and a half times the standard rate, applied to the whole request. The cheap flagship price, in other words, holds up to a quarter of the window; past that point, the meter runs faster.
OpenAI is also selling speed. New Ultrafast variants run the same models faster at much higher prices: Astra Ultrafast is available through Codex and the API at $40 input and $200 output per million tokens, and Sol Ultrafast through the API at $10 input and $50 output.
On OpenAI’s own benchmarks, the gains over the previous generation are broad. The company reports Astra scoring 73.8% on Terminal-Bench 2.0 against Sol’s 66.6%, 97.5% on GPQA Diamond against 95.8%, a perfect score on AIME 2026 against Sol’s 99.4%, and 84.2% on the CyberGym cybersecurity benchmark against 79.1%. Those are the maker’s numbers, measured by the maker.
The independent picture is more restrained. Artificial Analysis, which scores models on its Intelligence Index, puts Astra at 62, up from GPT-6 Astra’s 57, and Sol at 57.3, up from 55. That is a real improvement — and it still leaves Astra behind Claude Fable 5.2, which scores between 64.6 and 65 on the same index. OpenAI has cut its flagship’s price to match Anthropic’s; on the index the coverage cites, it has not yet caught Anthropic’s model.
The most consequential document in the release is not the price list but the system card, published on OpenAI’s Deployment Safety Hub. It states: “Both models meet the Critical threshold for Cybersecurity capabilities and the High threshold for Biological and Chemical capabilities under our Preparedness Framework. GPT-6.2 Astra meets the High threshold for AI Self-Improvement capabilities, while GPT-6.2 Sol falls below it.” OpenAI adds that Astra is “the first model we are deploying externally that we determine meets the High threshold” for self-improvement.
That phrase needs its context, and the card supplies it. In a change note, OpenAI clarifies that the High threshold for self-improvement is “an early milestone for proactively preparing safeguards for any models that may in future be capable of fully automated, recursive self-improvement.” High, in other words, does not mean Astra can rebuild itself. It means OpenAI judges the capability close enough that the safeguards for that future have to be prepared now, before a model arrives that can do it. The card rates only three categories — biological and chemical, cybersecurity, and AI self-improvement — and on the first two, both models sit at the same levels as the generation before them.
So the release, read whole, is two announcements in one. Commercially, frontier capability got cheaper: the flagship’s price fell to its chief rival’s, the smaller model followed it down, and speed is now a premium tier of its own. On safety, the same company’s own yardstick says its flagship has crossed a line it had not crossed in public before — not the line where a model improves itself without limit, but the line where its maker says preparations for that possibility must begin. Both halves are in OpenAI’s own documents. The prices are the part most users will feel this week; the rating is the part that will matter longer.
Catalina Clarke
AI News & Intelligence Desk, The Library Media
From The Library — the estate of our family.
Follow Us
