Skip to content

OpenAI releases GPT-6.2 Astra and Sol, Astra cheaper

Short answer

OpenAI released two models, GPT-6.2 Astra for hard projects and GPT-6.2 Sol for everyday work, both available in the API. Astra costs $8 per million input tokens and $40 output, below the previous Astra. Sol keeps its price. Early reviews of Astra are mixed.

What this means for operators

If your team runs agents for support triage, lead research or back-office tasks, the Astra price drop and the cheaper cached input matter most for long sessions that resend a lot of context. Before switching, rerun a fixed set of your own tickets or workflows on both models and compare cost per finished task. Check how Astra behaves in unattended runs, because early reviewers report it stopping to ask questions nobody could answer. Sol is the default candidate for routine volume work, since its price has not moved.

OpenAI has introduced two new models, GPT-6.2 Astra and GPT-6.2 Sol, according to a post by Simon Willison that quotes the announcement. OpenAI positions Astra as its frontier model for the hardest projects and Sol as the efficient, versatile model for daily work. Both are available through the API under the names gpt-6.2-astra and gpt-6.2-sol.

What the prices look like

Astra 6.2 is priced at 80% of the previous Astra: $8 per million input tokens and $40 per million output tokens, against $10 and $50 for 6 Astra. Willison notes that this matches the price of Claude Fable 5.2, which was also set at 80% of its predecessor's price.

Cached input gets a bigger cut. It is $0.40 per million tokens, where the older model charged $1. Willison calls this good news for longer agentic sessions, which tend to reuse the same context many times.

Sol stays at $2 input and $10 output per million tokens, the same as 6.1 Sol and 6 Sol. Willison points out that this is half the price of GPT-5.6 Sol, which was $4 and $20.

Early reception

Willison describes the buzz as mixed. According to him, Astra 6.2 does not score quite as high as Fable 5.2 on various benchmarks, and he has seen complaints that its communication style can be confusing.

He cites a review from Theo Browne, written as a feedback letter to OpenAI before release. Browne says the model handles many subagents on long tasks reasonably well, but then asks questions about internal details the user cannot answer. In his example, a task to build a new level for a game stalled two hours in because the model wanted to know which of two file names a reviewer step should write to.

Willison also ran his own animation prompt, a three.js scene of a city district turning into a giant mech, at the Ultra effort level in the Codex CLI. In his judgement neither new model did a good enough job on the mech transformation compared with Claude Fable 5.2 and Gemini 4 Carbon. This is one informal test by one reviewer, not a benchmark.

Willison also released LLM 0.38 with support for the new models, for anyone who tests models from the command line.

What to check

The price cut is confirmed in the source. The quality picture is still forming, and the reports above are individual impressions. Test on your own workloads before moving anything in production.

Source: Simon Willison

The tools in this story

  • OpenAI - price, who uses it, alternatives
  • Gemini - price, who uses it, alternatives
  • Claude - price, who uses it, alternatives
Next step

Visibility Analyzer

An AI news item will not tell you how assistants see your own site. The Visibility Analyzer checks that on your live site: which answers cite you, which pages an engine cannot retrieve, and what to fix first. Free to run.

Run a free visibility audit

Free to run. No card.

Fee
Free
Length
One run, minutes

Free tier: two analyses a day, no card required.