GPT-6.2 Astra and Sol Are Here: OpenAI's Frontier Model Gets 20% Cheaper

A close-up of a computer screen filled with colourful syntax-highlighted source code on a dark background

Photo: Hippopx (public domain). Illustrative image.

What happened. OpenAI launched GPT-6.2 Astra and GPT-6.2 Sol on 17 November 2026, its next generation of language models. Astra is the new frontier model for demanding professional work; Sol is the efficient, versatile model for everyday work. Both are available now in ChatGPT Work and Codex for Plus, Pro, Enterprise and Edu users, and in the API as gpt-6.2-astra and gpt-6.2-sol — but not in regular Chat. The pitch is familiar from the last two launches and sharper this time: better results on hard tasks with fewer output tokens, so users get more done inside the same usage limits.

The price is the headline. In the API, GPT-6.2 Astra costs $8 per million input tokens and $40 per million output tokens, with cached input at $0.40 — 20% below GPT-6 Astra’s $10/$50, and 60% cheaper on cached input. GPT-6.2 Sol keeps GPT-6.1 Sol’s pricing exactly: $2 input and $10 output per million tokens. Both models carry a 1.05-million-token context window. Pro 500 subscribers can use Ultrafast decoding with either model, and for the first time Ultrafast comes to Pro 200 subscribers too, starting with Sol. For the maths behind these per-token prices, see How Does AI API Pricing Work?.

The benchmarks are OpenAI’s own — and they are pointed. On Agents’ Last Exam, which tests multi-step professional work, OpenAI says Astra scores 5.2 points above GPT-6 Astra at less than a third of the cost, and 5.5 points ahead of Claude Fable 5.2 at just 5% of the cost. On Terminal-Bench-Science it scores 9 points higher than GPT-6 Astra at less than a third of the cost, and it sets a new high on the Vals Science Index. On legal work, Astra more than doubles Claude Fable 5.2’s pass rate on Harvey’s LAB-AA benchmark at about a sixth of the cost, with a dedicated Astra for Law configuration scoring higher still. Independent index Artificial Analysis ranks Astra second on its Intelligence Index at 62, behind Claude Fable 5.2 at 65, with Sol at 57.3. These are vendor-selected benchmarks and an independent index respectively — neither is a hands-on test, and OpenAI’s cost comparisons depend on its own per-task accounting.

Sol is the interesting model. On Cognition’s FrontierCode 2.0, which checks whether code changes meet the standards of real projects, GPT-6.2 Sol comes within 0.5 points of Astra at about a fifth of the cost. On LifeVal, OpenAI’s evaluation of challenging everyday tasks, Sol outscores GPT-6 Astra at about a seventh of the cost. The pattern matches the one Tencent’s Hy4 Flash set a day earlier: the efficient model is no longer a compromise for routine work, it is the rational default, with the frontier model reserved for the hardest jobs.

The safety section is the part to read twice. Both models meet the Critical threshold for cybersecurity and the High threshold for biological and chemical capabilities under OpenAI’s Preparedness Framework, and they launch with tighter safeguards for higher-risk cybersecurity help — protections OpenAI says also restrict some legitimate vulnerability research. Astra is the first externally released OpenAI model to meet the High threshold for AI self-improvement, which OpenAI defines as giving every one of its researchers the equivalent of an additional highly capable research engineer. In response it built Control Room, a system that tracks research workloads and can halt further model output when intervention is needed, plus expanded monitoring of models’ reasoning across internal workloads. The accompanying system card carries the full results.

Why it matters. GPT-6.2 is a price cut dressed as a capability launch: the frontier model got cheaper and better at the same time, and the cheap model got close enough to the frontier on code that the choice between them is now a cost decision, not a quality one. The self-improvement rating is the longer story — OpenAI is now shipping models it judges capable of accelerating AI research itself, and building the oversight machinery in the same release. Whether the safeguards hold is the question the next year of independent evaluation will answer.

Sources

  • GPT-6.2
  • GPT-6.2 Astra
  • GPT-6.2 Sol
  • OpenAI
  • AI models

Related articles