sysnote
News

GPT-6.2 Astra and Sol launch with tool calling restricted to the Responses API

OpenAI released GPT-6.2 Astra and GPT-6.2 Sol on November 17. Both accept Chat Completions requests, but applications that call tools have to use the Responses API. An existing agent built around Chat Completions needs an API migration as well as a new model name.

The models are rolling out in ChatGPT Work and Codex, and are available through the API as gpt-6.2-astra and gpt-6.2-sol. OpenAI positions Astra for demanding projects and Sol for everyday work. Its announcement reports that Sol comes within 0.5 points of Astra on Cognition's FrontierCode 2.0 at about one-fifth the cost per task, with both at maximum reasoning effort.

Old request defaults need changing

Both models reject reasoning effort none. The migration guide recommends starting new integrations at medium, or using low when replacing none or minimal. Applications using Chat Completions should omit temperature, top_p, logprobs and top_logprobs. Responses requests should omit temperature, top_p and top_logprobs, and remove message.output_text.logprobs from include.

These settings are worth checking in the wrapper that builds requests, where an old default can survive after the model ID has changed.

Editing the history can discard earlier reasoning

GPT-6.2 also changes how the API handles reasoning state. OpenAI says it drops reasoning items when it detects prefix edits, a model downgrade, or a change of end user or organization. A prefix edit changes conversation history that came before a reasoning item.

For an agent, that can happen during ordinary context housekeeping. Shortening an old tool result or rewriting a developer message changes the prefix. OpenAI warns that losing the affected reasoning can make the model repeat work, use more tokens or produce a worse answer. Its integration guide recommends appending updates and shortening large tool outputs before they first enter the conversation.

OpenAI's own compaction has an exception: /responses/compact and context_management preserve subsequent reasoning despite changing the prefix. A home-grown history rewrite should not be assumed to have that behavior. Responses expose input_transformations to show reasoning items removed and the reasons.

Prices and the usable context

The published Standard rates are per million tokens:

Model Input Cached input 30-minute cache write Output
GPT-6.2 Astra $8 $0.40 $10 $40
GPT-6.2 Sol $2 $0.10 $2.50 $10

The Astra and Sol model pages list a 1,050,000-token context window, with a maximum of 922,000 input tokens and 128,000 output tokens. Both take text and images and return text.

For requests above 272,000 input tokens, input and cache rates double and output rates rise by half for the full request. A 12-hour cache write also has its own higher rate: $20 per million tokens for Astra and $5 for Sol at short-context Standard pricing. The large context window is available, but an agent's growing history can move the entire request into a different price band.

Comments

No comments yet. Be the first.

Write a comment

Comments are published after review.