The most interesting line in this release isn’t a score. It’s a price comparison. OpenAI says GPT-6.2 Astra beats Claude Fable 5.2 on a long-horizon work benchmark at 5% of the cost per task. If that holds up outside OpenAI’s own charts, the frontier race has become an efficiency race.
OpenAI announced two models. Astra is the frontier option for hard projects. Sol is the cheaper, general-purpose model meant for daily use. Both are pitched as getting better results with fewer output tokens, which means more work inside your usage limits.
What Astra claims
On Agents’ Last Exam, which tests multistep professional work, Astra scores 5.2 points above GPT-6 Astra at under a third of the cost. It also finishes 5.5 points ahead of Claude Fable 5.2. OpenAI notes that Fable’s leaderboard results exclude nine tasks that used a fallback model, a detail that matters when you compare the two.
The science claims are equally bold. OpenAI calls Astra the best model in the world for scientists, citing new highs on Terminal-Bench-Science 0.1 (9 points above GPT-6 Astra at under one-third the cost) and the Vals Science Index. On professional documents, OpenAI says it trained Astra to follow templates and field conventions in law, finance, sales, and accounting, with less filler. On Harvey LAB-AA, a legal benchmark, it reportedly more than doubles Fable 5.2’s pass rate at about one-sixth the cost. There is also a law-specific configuration, Astra for Law, that scores higher still.
OpenAI’s own GDPval v2 shows Astra up 6.2 points over its predecessor while using 22% fewer output tokens. On AA-AnalystAgent, it answers 62.5% of questions correctly on all five attempts, which is a stricter test than a single lucky pass.
Sol is the practical one
For most teams, Sol matters more. OpenAI says it is smarter and more efficient than GPT-6.1 Sol, and that many OpenAI staff already use it as their default. On Cognition’s FrontierCode 2.0, Sol lands within 0.5 points of Astra at roughly one-fifth the cost. On LifeVal, a benchmark of everyday tasks like picking kindergartens or planning meals, it beats GPT-6 Astra at about one-seventh the cost.
That is the real pitch. A model nearly as good as the flagship, at a fraction of the price, is what most production workloads need. Anthropic and Google have pushed similar tiering, but OpenAI is leaning harder on cost per task as the headline metric.
Pricing and availability
Astra’s API prices drop 20% from GPT-6 Astra. Cached input gets a 60% discount. Here is the full rate card, per million tokens:
- Input: $8
- Cached input: $0.40
- Output: $40
Sol keeps GPT-6.1 Sol’s prices, so its gains come free. Both models are available in ChatGPT Work.
Should you care?
Yes, with caveats. Nearly every number here is from OpenAI or from benchmarks it chose to feature, and some, like GDPval v2, are its own. Cost-per-task comparisons also depend on how many tokens each model burns, which varies by harness. Claude Fable 5.2 was tested in a mini-swe-agent harness on some science tests, which isn’t an identical setup.
Still, the direction is clear. Quality gaps between top labs are narrow, so price and token efficiency now decide who wins contracts. If you run agents at volume, test Sol against your current default before anything else. If your work is legal, financial, or scientific, Astra deserves a bake-off against Fable 5.2 on your own data. Trust your evals over anyone’s chart, including this one.



