A frontier model that costs less than the one it replaces
OpenAI released GPT-6.2 Astra and GPT-6.2 Sol on 17 November, and the number that stands out is not a benchmark but a price. Astra, the frontier model, now costs $8 per million input tokens, $0.40 per million cached input tokens and $40 per million output tokens. That is a 20% cut on GPT-6 Astra’s rates, with a deeper 60% discount on cached input. Sol, the cheaper model aimed at everyday work, keeps GPT-6.1 Sol’s prices of $2, $0.10 and $10.
OpenAI said both models also reach the same result using fewer output tokens, so the cost of a finished task falls further than the per-token figures alone suggest. On its internal GDPval v2 benchmark, the company said Astra scored 6.2 percentage points above GPT-6 Astra while using 22% fewer output tokens, and that Sol improved 5.3 points on GPT-6.1 Sol.
The numbers are OpenAI’s own
Every figure in the announcement comes from OpenAI, and most are comparisons with named rivals. On Agents’ Last Exam, a test of long-horizon professional work, OpenAI said Astra scored 5.2 points above GPT-6 Astra at less than a third of the cost, and 5.5 points above Anthropic’s Claude Fable 5.2 at 5% of the cost. The company noted that the leaderboard results it cited for Fable 5.2 exclude nine tasks that used a fallback model.

On Harvey’s LAB-AA v1.1 legal benchmark, OpenAI said Astra more than doubled Claude Fable 5.2’s pass rate at about a sixth of the cost. On Artificial Analysis’s AA-AnalystAgent, which asks models to pull figures out of real spreadsheets and reports, it said Astra answered 62.5% of questions correctly on all five independent attempts. On AA-3D, where human judges compare two blinded browser experiences built from the same prompt, OpenAI reported Astra first at 1,318 Elo, Claude Fable 5.2 at 1,271 and Sol at 1,259.
Sol’s case is that it is close enough to Astra. In Cognition’s FrontierCode 2.0, OpenAI said Sol came within 0.5 points of Astra at about a fifth of the cost, with both models at maximum reasoning.
Slides, spreadsheets and CAD
The pitch is professional output rather than conversation. OpenAI said it trained Astra to follow the conventions of fields such as law, finance, sales and accounting, to match a slide template’s layout, colours and typography from a file or even a screenshot, and to build valuation workbooks that reconcile three financial statements. It also said the model’s computer use is precise enough to work on individual parts and connections inside CAD and simulation tools.

More cyber capability, tighter limits
Both models meet the Critical threshold for cybersecurity and the High threshold for biological and chemical capabilities under OpenAI’s Preparedness Framework. The company said it is launching them with tighter safeguards for higher-risk cybersecurity assistance, and acknowledged that those protections also restrict some legitimate vulnerability research and security testing. Defenders who need fewer refusals are pointed at Daybreak, which currently offers GPT-6 Astra rather than the new models.
Availability
Both models are available from 17 November to Plus, Pro, Enterprise and Edu users in ChatGPT Work and Codex, and to developers through the API as gpt-6.2-astra and gpt-6.2-sol. Pro 500 subscribers get OpenAI’s Ultrafast mode with either model, and Pro 200 subscribers get it for the first time, starting with Sol. Enterprise and Edu access depends on a workspace administrator enabling it.
What to watch is whether independent evaluators reproduce the cost-per-task claims. OpenAI’s own figures put Astra ahead of Claude Fable 5.2 at a twentieth of the price on one benchmark. That is the kind of gap outside testers move quickly to check.