GPT-6.2 Sol is the cheap one to try; Astra's reviews are rough
Astra ranks second on Artificial Analysis, but hands-on testers were unimpressed.
OpenAI shipped two models on the same day. GPT Astra is the expensive frontier one, second on the Artificial Analysis index. GPT Sol is the everyday one, unchanged at $2 in and $10 out per million tokens.
The deeper look
The benchmarks look good. Artificial Analysis ranks Astra second, at 62 against 65 for Claude Fable 5.2, for far less per task. OpenAI itself claims a 5.5-point lead over Fable on Agents' Last Exam at about 5% of the cost, which is the maker's word. One Hacker News commenter reports Astra five points below Fable 5.2 on FrontierBench, so the sources disagree.
Hands-on, the picture sours. Matt Shumer called Astra and Sol essentially unusable, with his Opus and Fable agents rejecting much of the subagent work. Theo Browne found it handled many subagents but asked incomprehensible internal-detail questions that blocked a two-hour run. AI Search found no clear gain over the previous Astra.
So Sol is the more interesting buy: the same price as before, and the one to test first. Astra's price is down too, to 80% of the old $10/$50.
It is for builders comparing agent costs. Skip Astra if you need low hallucination rates, and test Sol on your own tasks first.
How to start: call gpt-6.2-sol or gpt-6.2-astra through the API, or see openai.com/index/introducing-gpt-6-2-astra-sol.
$ curl https://openrouter.ai/api/v1/chat/completions -H "Authorization: Bearer $OPENROUTER_API_KEY" -H "Content-Type: application/json" -d '{"model": "openai/gpt-6.2-astra", "messages": [{"role": "user", "content": "Hello"}]}'