AI Model Report

Reviews · NOVEMBER 18, 2026

GPT-6.2 Astra cuts token prices; Sol closes in on one coding test

OpenAI’s new models warrant a controlled comparison on customer-facing work, but the launch benchmarks do not establish better copy, more leads, or higher sales.

By Karl Strauchman · Senior model reviewer · November 18, 2026

OpenAI released GPT-6.2 Astra and GPT-6.2 Sol on November 17, cutting Astra’s standard API input and output token prices by 20% against GPT-6 Astra, according to its announcement and API changelog. OpenAI also reports that Sol came within 0.5 points of Astra on Cognition’s FrontierCode 2.0 at about one-fifth the cost.

Both models are available through the API, ChatGPT Work, and Codex. OpenAI positions Astra for demanding projects and Sol for everyday work, and says both can complete demanding tasks with fewer output tokens. Those claims warrant testing, but the coding result doesn’t establish better copywriting or higher sales.

Token rates and migration details

OpenAI lists Astra at $8 per million input tokens, $0.40 per million cached input tokens, and $40 per million output tokens. Sol retains GPT-6.1 Sol’s rates. The API changelog lists that predecessor’s standard rates for prompts up to 272K input tokens as $2 input, $0.10 cached input, and $10 output per million tokens.

For an operator choosing between them, the distinction is between the listed token rates and the cost of completing a task. OpenAI’s claim of fewer output tokens suggests another possible source of savings, but a small team would need to measure that on its own work.

The November 17 changelog says both models support Responses and Chat Completions; tool calling requires Responses. It also flags changed handling of conversation-prefix edits and top-level instructions. That makes migration testing relevant beyond a comparison of token bills: teams should check how existing conversations and instructions behave before delegating customer-facing work.

What the coding comparison does, and doesn’t, show

FrontierCode 2.0 evaluates whether changes to real codebases are suitable for maintainers to merge. According to 9to5Mac’s account of OpenAI’s announcement, both models used maximum reasoning effort in the comparison. Sol’s proximity to Astra is therefore a self-reported result on a coding test under that setting, not evidence of equivalent performance across tasks.

The benchmark doesn’t measure whether a proposal is accurate, website copy is usable, or sales material brings in customers.

The benchmark doesn’t measure whether a proposal is accurate, website copy is usable, or sales material brings in customers.

For an owner or founder, a useful comparison would test the same customer-facing assignments on both models and track total cost per approved asset, including human review and revisions. Lower API rates may make those tests more practical; they don’t establish more leads or sales.

Scope adherence deserves a separate check. Quartz reports that OpenAI scrapped an earlier GPT-6.1 Astra update over concerns about staying within scope and communicating completed work. That reporting supports a concrete test before delegation: whether each model follows the assigned boundaries and accurately describes what it completed, not merely whether its first draft reads well.

Sources