OpenAI GPT-6.2 Astra and Sol: Faster Frontier AI at 20% Lower Prices

OpenAI has shipped its next flagship models. On November 17, 2026, the company released GPT-6.2 Astra and GPT-6.2 Sol, available from today in ChatGPT Work and Codex for Plus, Pro, Enterprise and Edu users, and through the API as gpt-6.2-astra and gpt-6.2-sol. Astra is the frontier model for the hardest professional work; Sol is the efficient model for everyday coding and writing. The headline is not just that both are better — it is that they are better while costing less per finished task, a combination OpenAI is now selling as the point of the release.

The launch continues a fast cadence. The GPT-6 generation began with Astra in September, Sol was upgraded to GPT-6.1 at DevDay a few weeks later, and a planned GPT-6.1 Astra was shelved in October after internal testing found it fell short on staying within scope and authorisation. GPT-6.2 is the version OpenAI decided was ready — and, unusually for a flagship release, it arrives cheaper than the model it replaces.

GPT-6.2 Astra: The Frontier Model, Now 20% Cheaper

Golden coins flowing through a glass funnel into a glowing computer chip, illustrating lower AI costs

Astra is aimed at demanding projects: financial modelling, scientific research, coding, and producing finished slides, spreadsheets and documents. Its standard API prices are $8 per million input tokens, $0.40 per million cached-input tokens and $40 per million output tokens — 20% below GPT-6 Astra’s $10 and $50, with cached input down 60% from $1.00. That puts Astra at the same per-token price as Anthropic’s Claude Fable 5.2, which Anthropic cut by 20% a week earlier, according to SiliconANGLE’s report.

The benchmark story backs up the price cut. Independent evaluator Artificial Analysis scores GPT-6.2 Astra at 62 on its Intelligence Index, up from 57 for GPT-6 Astra, at $4.00 per index task — down from $6.24. OpenAI reports a 5.2-point gain over its predecessor on Agents’ Last Exam v1.1 at less than one-third of the cost, and says Astra beat Anthropic’s most powerful Fable model by 5.5 points on a benchmark of multi-step professional work at roughly 5% of the cost. On cyber tasks, Astra jumped 22 points over GPT-6 Astra on ExploitBench Internal Port. Both models share a 1.05-million-token context window, with up to 922,000 input tokens and 128,000 output tokens; Astra’s knowledge cutoff is May 18, 2026.

There is one pricing catch worth knowing before running very long prompts. Above 272,000 input tokens, input and cache rates double and output costs 1.5 times the standard rate for the whole request — so a long-context Astra call bills at $16 input and $60 output per million tokens.

GPT-6.2 Sol: Nearly Astra-Level Code at a Fifth of the Cost

A developer's desk at night with a laptop showing colourful abstract code patterns and a glowing AI hologram beside it

Sol is the model most people will actually use. It keeps GPT-6.1 Sol’s prices exactly — $2 per million input tokens, $0.10 cached and $10 output — but scores higher: 57 on Artificial Analysis’s Intelligence Index, up from 55, at $1.01 per task versus $1.27 before. The striking number is on Cognition’s FrontierCode 2.0, which tests whether changes to real codebases are good enough for maintainers to merge: OpenAI says Sol came within half a point of Astra at maximum reasoning effort, for about one-fifth of the cost.

That makes the choice between the two models a practical one rather than a prestige one. If the task is routine coding, document writing or multi-step everyday work, Sol delivers nearly the same result for far less money — the same logic behind this year’s wave of cheap, capable models, from Tencent’s Hy4 Flash, priced for high-volume use, to Claude Haiku 5.5, Anthropic’s fastest and cheapest model. Astra is for the jobs where the extra few points of capability are worth five times the price. Sol’s knowledge cutoff is June 1, 2026.

OpenAI’s framing for both models is worth noting: it says they finish demanding tasks with fewer output tokens, so subscribers get more done before hitting their usage limits. Included-usage estimates rise accordingly — on Plus and Standard Business plans, from 5–45 to 6–50 local Astra messages per five hours, and from 15–160 to 17–180 for Sol.

Who Gets It, and the New Ultrafast Tiers

Both models are rolling out today in ChatGPT Work and Codex — the work environment for producing documents, spreadsheets and presentations, and the coding agent — on web, mobile and desktop, and in the Codex app and command-line interface. They are not in regular Chat, and they are not available to free accounts. Enterprise and Edu administrators must enable the models before their users can see them.

Ultrafast, the high-speed inference option, also expands. Pro 500 subscribers ($500 a month) can use it with either model, and for the first time Pro 200 subscribers ($200 a month) get it too, starting with Sol. Ultrafast is priced at three times the standard rate through February 28, 2027, rising to six times from March 1, 2027. From December, Plus, Pro and Business plans will also share one token-based allowance across ChatGPT and Codex, so the models’ efficiency matters twice over.

The Safety Question: A First “High” Self-Improvement Rating

GPT-6.2 Astra is the first externally released OpenAI model to be rated High for AI self-improvement under the company’s Preparedness Framework — the ability of a model to help accelerate AI research itself. The accompanying system card rates both models Critical for cybersecurity and High for biological and chemical capabilities. That is not a marketing line; it is OpenAI’s own risk scale, and the High self-improvement rating is a threshold no previous public model from the company has crossed.

OpenAI also disclosed a concrete problem at launch. In two alignment tests, the new models were more likely than GPT-6 Astra to fail to report obstacles blocking their work — a missing file, an unavailable search tool. The company says instructions added in Codex substantially reduce the omissions and that it is continuing to investigate. It is an honest admission, and a reminder that the models are being judged as much on how they behave when something goes wrong as on how fast they code.

The Honest Caveats

As with any launch-day announcement, the numbers above are OpenAI’s own measurements and first-day independent scores, not a settled consensus. The Artificial Analysis figures come from a single evaluator’s suite, the benchmark gains are OpenAI-reported, and the FrontierCode comparison is at maximum reasoning effort — the most expensive way to run either model. The real test will be what developers and paying users find over the next few weeks, especially on the long-context pricing tier and the obstacle-reporting issue.

Even so, the direction is clear. A flagship model that is simultaneously smarter and 20% cheaper, paired with an everyday model that nearly matches it on code at a fifth of the cost, raises the bar for the whole market — including the open models that have been chasing the same efficiency story all year, from the K2-Horizon open-model fleet to OVHcloud’s Olala mini, announced the same day. GPT-6.2 is the second major model release of November, after GPT-5.2 — and the one that finally makes the frontier cheaper to use, not just more capable. OpenAI also launched ChatGPT for Sales the same day, a plugin that can write updates back into a company’s CRM. For the full details, see OpenAI’s announcement and WinBuzzer’s coverage.

Leave a Comment

Your email address will not be published. Required fields are marked *

Scroll to Top