· via dev.to (home feed)
OpenAI ships GPT-6.2 Astra and Sol with cheaper tokens and mixed early reviews
OpenAI's new GPT-6.2 Astra and Sol models are live via API, with 20% cheaper pricing, a 60% cached-input discount, and mixed early reviews.

Two new GPT-6.2 models, live now
OpenAI added two models to the GPT-6.2 family on 17 November 2026: Astra, a flagship model built for the hardest tasks, and Sol, a cheaper everyday model. Both are live via API now as gpt-6.2-astra and gpt-6.2-sol, and the pair arrived with unusually fast ecosystem support — Simon Willison has already shipped LLM 0.38, adding support for both models to his popular command-line tool.
According to OpenAI, both models produce stronger results on demanding tasks while generating fewer output tokens, which in practice should let more work fit inside the same subscription and token budgets.
Astra's benchmark claims
On agents, OpenAI says Astra scores 5.2 points above GPT-6 Astra on Agents' Last Exam, at under a third of the cost. Against Claude Fable 5.2, Astra leads by 5.5 points at roughly 5% of the cost. The model also sets new highs on Terminal-Bench-Science 0.1 and the Vals Science Index. OpenAI is calling Astra the best model in the world for scientists, per a post on dev.to.
Pricing: 20% cheaper, plus a cache discount
Astra costs $8 for input and $40 for output per million tokens, about 20% cheaper than GPT-6 Astra, per pricing OpenAI published and Willison confirmed. Cached input drops to $0.40 per million tokens after a 60% discount. Sol keeps GPT-6.1 pricing at $2/$10.
For agent builders, the cache discount may matter more than the headline price cut. Per Willison, the steep discount on cached input should help developers running long agent sessions, since the same context would otherwise be billed again and again as a session grows.
Sol: the everyday model with near-flagship scores
OpenAI says many of its own employees use Sol as their default model for both technical and non-technical work. On Cognition FrontierCode 2.0, Sol trails Astra by only 0.5 points at roughly a fifth of the cost, per OpenAI's data. On LifeVal it beats GPT-6 Astra at about a seventh the cost.
Mixed reviews from early users
The independent verdict so far is mixed. Willison notes that Astra does not consistently score above Claude Fable 5.2 across benchmarks. Users complain about a confusing communication style. Theo Browne reports the model asks odd questions about internal implementation when working with subagents.
In Willison's own animation-prompt test, both GPT-6.2 models trail Fable 5.2 and Gemini 4 Carbon. Willison also flags a caveat worth keeping in mind when comparing models directly: the Fable 5.2 leaderboard result excludes nine tasks where a fallback model was used.
Why it matters
For AI developers, the practical takeaway is straightforward: Astra gets cheaper and stronger on agent workloads, while Sol offers near-flagship performance at a small fraction of the price. The 60% discount on cached input is arguably the biggest win here, since it makes long-running agent sessions dramatically more affordable. But the release is not a clean sweep — mixed reviews and benchmark caveats suggest a closer race with rivals than OpenAI's own numbers imply. Developers building on the new models should treat the benchmark claims as a starting point, not a conclusion, and test the pair against their own workloads before committing.
- #openai
- #gpt-6
- #ai-models
- #ai-agents
- #ai-benchmarks
- #ai-api