OpenAI’s newest flagship model costs a fifth less to run than the one it replaces, and the company says it now beats Anthropic’s Claude Fable 5.2 on the hardest benchmark it has.

On November 17, OpenAI shipped GPT-6.2 Astra and GPT-6.2 Sol to ChatGPT Work, Codex and the API. Astra is pitched at the hardest professional and scientific work. Sol is the cheaper, faster model meant for everyday tasks. Both replace the GPT-6 generation OpenAI launched back in September.
The pricing move is the headline for anyone actually paying the bill. Astra’s standard API rate drops to $8 per million input tokens and $40 per million output tokens, a 20 percent cut from GPT-6 Astra. Cached input is down 60 percent. Sol keeps the $2 input and $10 output pricing it already had under GPT-6.1.
OpenAI says the real gain isn’t just the sticker price, it’s that both models finish tasks using fewer output tokens than their predecessors, which compounds the savings on top of the rate cut. On the company’s GDPval v2 benchmark, built from real consulting, finance and design assignments, Astra improved 6.2 percentage points over GPT-6 Astra while using 22 percent fewer output tokens. Sol improved 5.3 points over GPT-6.1 Sol.
The comparison that will get the most attention in AI circles is against Anthropic. OpenAI says GPT-6.2 Astra more than doubles Claude Fable 5.2’s pass rate on an internal law-focused evaluation, using about one-sixth the cost per task. On Artificial Analysis’s AA-3D benchmark, where human judges pick between two blinded experiences built from the same prompt, OpenAI reports Astra ranking first at 1,318 Elo, ahead of Claude Fable 5.2 at 1,271. Independent testers at Decrypt reached a more mixed verdict, ranking Astra second behind Fable 5.2 overall and finding no real improvement on coding-agent tasks specifically. Benchmark results from the company that built the model are worth treating as a claim, not a verdict, until outside labs run their own numbers.
There is a notable gap in the release timeline worth flagging. OpenAI had planned to ship GPT-6.1 Astra back in October, then called it off after internal testing found the model fell short on staying within scope and authorization, meaning it occasionally did things beyond what it was actually asked to do. That is not a small thing to shelve a flagship release over, and it is worth remembering the next time a lab promises a model soon.
GPT-6.2 Astra also comes with a new classification attached. It is the first OpenAI model released externally that meets the company’s High threshold for AI self-improvement under its Preparedness Framework, a measure of how much a model can accelerate AI research itself. Both Astra and Sol meet the Critical threshold for cybersecurity risk, the same tier GPT-6 Astra sat at. OpenAI says it is launching both with tighter safeguards around higher-risk cybersecurity use because of that jump in capability.
Early access partners are already reporting practical gains. Legal AI firm Harvey, contract platform Legora and project management tool ClickUp all told OpenAI they are seeing fewer tokens burned and fewer review rounds needed per task. Accounting automation company Basis says long-horizon bookkeeping work is now taking roughly half the time it used to.
One of the more immediately useful upgrades for non-technical users: Astra can now take a screenshot of an existing slide and replicate its layout, colors and font sizing for the rest of a deck. For anyone who has ever tried to match a client’s exact brand template by eye, that is a real time-saver, assuming it actually works as described once more people get hands-on with it.
Access rolled out immediately to Plus, Pro, Enterprise and Edu users inside ChatGPT Work and Codex, with developers able to call gpt-6.2-astra and gpt-6.2-sol through the API the same day. Pro 500 subscribers get the fastest Ultrafast tier with either model at launch, while Pro 200 subscribers get it first paired with Sol only.
The bottom line
A 20 percent price cut on a frontier model paired with real efficiency gains is a genuine win for anyone building on the API, and the early partner numbers from Harvey and Basis suggest it is not just a benchmark story. But the shelved October release and the new self-improvement classification are the parts of this announcement worth sitting with longer than the Elo scores. Faster and cheaper is good news. Faster, cheaper and harder to keep inside its intended scope is the thing worth actually watching as adoption spreads.



