OpenAI's GPT-6.2 Astra ties Claude Fable 5.2 on Epoch's index, but trails it on real tasks
Astra is 20% cheaper and sets a math record, but Epoch says it left out estimates a contractor couldn't.
On November 17, 2026, OpenAI released GPT-6.2 Astra and GPT-6.2 Sol, new versions of its frontier model and its everyday one. Both went live that afternoon in ChatGPT Work, Codex and the API for Plus, Pro, Enterprise and Edu users.
Two outside benchmark groups posted results within an hour. Both put Astra at or near the top of their main index, and Epoch AI's tests on its own research work put it well behind Anthropic's best.
- Astra API price
- $8/$40 per million tokens, 20% below GPT-6 Astra
- Sol API price
- unchanged from GPT-6.1 Sol, at $2/$10
- Artificial Analysis Intelligence Index
- second, behind Claude Fable 5.2
- Epoch Capabilities Index
- 174, on par with Fable 5.2
- FrontierMath Erdős
- 7/68 solved, previous record 3/68
What OpenAI is claiming
OpenAI's launch thread led with cost, and it named the kinds of work where it says you'll see the change.
GPT-6.2 Astra—a frontier model built for your hardest projects. It sets new highs on evals across professional work and scientific research at less than a third of the cost per task of GPT‑6 Astra. You’ll see the difference in things like slideshows, websites, and 3D designs.
The cyber post in the same thread is about what OpenAI is holding back. It calls Astra its most capable model ever for cyber, 22 points ahead of GPT-6 Astra on ExploitBench Internal Port, so both new models launch with tighter safeguards on higher-risk security help. OpenAI says it's working to bring them to Daybreak, where approved defenders would get them with reduced refusals.
Sam Altman's own post picked out the cheaper model. He called both very token-efficient and said 6.2 Sol ultrafast is "something I think people will really enjoy." Pro 500 subscribers get Ultrafast on either model, and Pro 200 subscribers get it for the first time, starting with Sol.
What the outside tests found
Artificial Analysis puts Astra second on its Intelligence Index, behind only Claude Fable 5.2. That's a 5-point jump over GPT-6 Astra, enough to pass Gemini 4 Carbon (60) and Claude Opus 5.5 (59). Sol gains 2 points to 57, level with the old Astra at about 15% of the cost per task.
GPT-6.2 Astra is flat with GPT-6 Astra, and GPT-6.2 Sol loses 3 points relative to GPT-6.1 Sol in the Artificial Analysis Coding Agent Index
Coding is where it stalls. Sol's drop comes mostly from one test, where it's more likely than its predecessor to leave key facts out of its answers. Both models also guess a lot when they don't know a factual answer: Artificial Analysis measures hallucination rates of 59% and 55% at max effort.
Epoch AI scores Astra 174 on its capabilities index, level with Fable 5.2 and above what the pre-Astra trend for reasoning models predicted. Its record is in mathematics. On FrontierMath Erdős, problems that were unsolved as of August 2026, Astra produced Lean-verified solutions to 7 of 68 on a budget of about $240 a question, one of them a problem with no known solution before. The previous best, GPT-6.1 Sol's, was 3.
Then Epoch ran it on realistic tasks drawn from its own work. There Astra fell behind Fable 5.2 and barely improved on GPT-6 Astra. Epoch says it missed the core question on simple research projects, and on a data center research task it left out critical estimates, which Epoch says "would be unacceptable for a contractor."
Ethan Mollick, who tested the model before launch, read that gap as a case for judging models by hand as well as by score.
The reviews of the early testers like myself of GPT 6.2 Astra is a good vindication of the continuing value of vibe benchmarks in addition to numeric ones. We lack good measures of personality or perseverance or creativity or writing clarity or taste. But these things matter
What it means if you pay for it
For a business, the price list is the concrete part. Astra costs $8 per million input tokens and $40 per million output in the API, 20% below GPT-6 Astra, and cached input drops to $0.40 from $1.00. Sol stays at $2 and $10. A company that runs a model over documents all day gets a cheaper frontier tier without switching vendors.
The catch is the one Epoch wrote down. The work people actually hand these models, a research memo or an estimate for a client, looks more like Epoch's own tasks than like a leaderboard, and on those Astra still trails Fable 5.2. I'd also take the hallucination number seriously: at 59%, Astra guesses more often than not when it doesn't know, so anything it writes for you still needs checking.
Epoch says it will publish a full report on the updated version of its Automation Reports benchmark in another post.
More on OpenAI
- Calif says one researcher using OpenAI's Codex found a critical Java flaw in an hourNovember 18, 2026
- OpenAI brings four outside evaluators inside, but one says the access isn't employee-likeNovember 17, 2026
- California sues OpenAI, saying the company lost control of its AI agents at least 80 timesNovember 16, 2026
- ChatGPT Wallet lets OpenAI's dots buy things for you, but you approve each purchaseNovember 14, 2026
