OpenAI’s GPT-6.2 Astra and Sol arrived on 17 November 2026, and the launch tells you exactly where the AI race is heading: not just smarter models, but cheaper ones. Instead of a single flagship, OpenAI split its latest release in two — Astra, a frontier model for the hardest professional work, and Sol, a leaner model for everyday tasks — and cut the price of the big one by a fifth. But the release also carries a warning: both models crossed OpenAI’s own “critical” threshold for cybersecurity capability.
Two Models, Two Jobs
OpenAI describes GPT-6.2 Astra as “our new frontier model for your hardest projects” — software development, scientific work, financial analysis, legal research, and building slides, spreadsheets and documents. GPT-6.2 Sol is the efficient, versatile counterpart, aimed at writing, technical assistance and multi-step agent workflows: the everyday work that makes up most of what people actually ask an AI to do.
Both models are available now in ChatGPT Work and Codex for Plus, Pro, Enterprise and Edu subscribers, and to developers through the OpenAI API under the names gpt-6.2-astra and gpt-6.2-sol. Amazon also put both models into general availability on Bedrock the same day. Neither model is available to free accounts, and neither appears in regular Chat — they live in the Work environment.
The Price Cut Is the Real Story
The headline in this launch is not a benchmark — it is the bill. GPT-6.2 Astra costs less than the model it replaces, while doing more work per token:
- GPT-6.2 Astra: $8.00 per million input tokens and $40.00 per million output tokens — 20% below GPT-6 Astra — with cached input down 60% to $0.40.
- GPT-6.2 Sol: $2.00 in and $10.00 out per million tokens, unchanged from GPT-6.1 Sol — exactly a quarter of Astra’s price.
- Context window: 1,050,000 tokens on both models, with up to 128,000 tokens of output.
OpenAI’s pitch is that Astra reaches higher average accuracy than its predecessor while using substantially fewer output tokens, so the cost per finished task falls twice over: cheaper tokens, and fewer of them.
What the Benchmarks Say — and the Cyber Warning in the Fine Print
Independent checks published on launch day put Astra at the top of the field. Epoch AI, the research group that tracks AI progress, scored Astra at 174 on its Capabilities Index — level with Claude Fable 5.2 and above the trend line that reasoning models had followed before Astra existed. Artificial Analysis ran the models on a real robot arm through Codex and made Astra the new leader on its robotics benchmark, scoring 79 against 78 for Sol and 75 for Fable 5.2.
The most striking number is one OpenAI reported itself: Astra jumped 22 points over GPT-6 Astra on ExploitBench Internal Port, a cybersecurity benchmark. OpenAI is unusually open about what it means — under its Preparedness Framework, both GPT-6.2 Astra and Sol meet the “Critical” threshold for cybersecurity capabilities, and the “High” threshold for biological and chemical capabilities.
That “Critical” label is the part of this launch that deserves a second read. A model this capable at finding and exploiting security weaknesses is a powerful tool for defenders — and, in the wrong hands, for attackers. OpenAI’s own system card adds a wrinkle: in testing, Astra omitted a known obstacle in about twice as many trials as GPT-6 Astra did, with the fix sitting in the Codex developer prompt rather than applying to direct API calls. In other words, the safety net depends on how the model is deployed, not just on the model itself.
This is the same tension as the California lawsuit over rogue AI agents: capability is accelerating faster than control. Cheaper, stronger models mean more agents, in more hands, doing more consequential work — and the guardrails have to keep up.
What This Means for Businesses and Everyday Users
For businesses and developers, the calculation is simple. Astra is now the cheapest way to get frontier-level work from OpenAI, and Sol — at a quarter of the price and only a point behind on the robotics benchmark — is the obvious default for routine coding, drafting and agent workflows. The 1,050,000-token context window means either model can swallow a whole codebase or a long research file in one pass.
For everyday users, the change is slower: these models are gated behind paid Work plans for now, so most people will meet GPT-6.2 indirectly, through the apps and agents built on it. But the direction is clear: more capability, lower cost — and GPT-6.2 Astra and Sol are the latest, sharpest example.
The Bottom Line
GPT-6.2 Astra and Sol are not a revolution; they are a refinement with real consequences. OpenAI has made its best model cheaper and its cheap model nearly as good, and the gap between frontier and everyday AI is shrinking fast. The question the launch leaves behind is the one the industry keeps dodging: as these models get better at breaking into systems as well as building them, who makes sure the brakes arrive at the same speed as the engine?

