GPT-6.2 Astra cuts API prices by 20% and ranks second on Artificial Analysis, but its coding agent score stayed flat
OpenAI shipped GPT-6.2 Astra and GPT-6.2 Sol, and the price changes matter more to automation builders than the headline benchmarks.
What happened
OpenAI released two models on November 17: GPT-6.2 Astra for demanding work and GPT-6.2 Sol for everyday tasks. In the API, Astra costs $8 per million input tokens, $0.40 for cached input and $40 for output, which The Decoder reports is 20% below its predecessor. Simon Willison notes the cached price fell from $1 to $0.40, a large cut for long agent sessions. Sol stays at $2 input, $0.10 cached and $10 output.
On the Artificial Analysis Intelligence Index, Astra scores 62 and ranks second behind Claude Fable 5.2 at 65. Sol reaches 57, matching GPT-6 Astra at roughly 15% of the cost per task. OpenAI also says Astra gained 6.2 points on its GDPval-v2 benchmark while using 22% fewer output tokens.
The independent results are mixed. Artificial Analysis found Astra's Coding Agent Index score unchanged from GPT-6 Astra, and Sol dropped three points, mostly because it more often omits key facts on one question answering test. At maximum effort, Astra guessed instead of admitting uncertainty on 59% of a set of very hard factual questions, and Sol on 55%. Willison also quotes Theo Browne, who described Astra stopping two hours into a long task to ask him an internal implementation question he had no way to answer.
My take
For most business automations, Sol is the model to look at. Classification, extraction and routing steps run thousands of times a day, and a $0.10 cached input price makes long, reused system prompts cheap.
I wouldn't change a model ID in production on the strength of a launch post, though. The omitted facts regression is the kind of change that breaks an extraction step without throwing an error: the JSON still parses, a field is simply empty. Run your existing eval set against the new model, compare field by field, and keep the old ID pinned until it passes.
The Theo Browne example points to a design rule for long agent runs. Decide in advance what the agent may ask a human, set a timeout on any question it raises, and give it a default action so a two hour job doesn't sit blocked on a question nobody can answer.
More posts
- LangSmith can now hide trace inputs and outputs from engineers while still showing them latency, cost and run treesNov 19, 2026
- Microsoft-Decision-1.1 is generally available for routing and classification at $0.042 per million input tokens, with free outputNov 19, 2026
- LangSmith now ties coding agent spend to pull requests, so teams can see cost per merged PR across Claude Code, Codex and CursorNov 19, 2026
