OpenAI shipped GPT-6.2 Astra on November 17, 2026, alongside its cheaper sibling GPT-6.2 Sol. The headline specs look attractive: a 1,050,000 token context window, $8 per million input tokens (down from $10 for the previous Astra), and $40 per million output tokens. But developers need to read the fine print before swapping their model IDs. The Responses API is now the primary endpoint for new features, and while Chat Completions is supported, it lacks access to hosted tools like web search and computer use.

The Reasoning Effort Tax

The single biggest cost driver is reasoning effort. OpenAI confirms that 'none' is rejected for both 6.2 models, forcing every request to incur reasoning tokens. Early benchmarks from Simon Willison show that moving from 'high' to 'max' effort on the same image prompt costs 88.71 cents versus 11.55 cents—an eightfold increase. While per-token pricing dropped 20% compared to GPT-6 Astra, 'max' effort reasons longer, potentially making complex tasks more expensive than before. Always set effort explicitly; the default is undocumented and risky for reproducibility.

The 272K Token Cliff

Long-context pricing hides a brutal penalty. If your request exceeds 272,000 input tokens, the input rate doubles to $16 per million and output jumps to $60 per million for the entire request, not just the overflow. Feeding whole repositories into Astra without chunking or caching will drain your budget fast. Prompt caching helps—cached input drops to $0.40 per million—but you must structure your system prompts and tool lists carefully to stay under the cliff edge.

Migration Gotchas

Codebases migrating from GPT-6 Astra need more than a string swap. Any code path sending 'effort': 'none' will fail immediately. Hosted tools like MCP, code interpreter, and apply patch are only available on the Responses endpoint. Early reviews are mixed: Ethan Mollick called the model 'very irregular and weird,' and Dan Shipper noted it often stops early in end-to-end builds. Test your specific prompts before deploying to production.

Key Takeaways

  • Input costs $8/M tokens; output costs $40/M tokens. Cached input is $0.40/M.
  • Requests over 272K input tokens trigger a 2x input and 1.5x output penalty.
  • 'Effort: none' is rejected; 'max' effort can cost 8x more than 'high'.
  • Hosted tools (web search, computer use) require the Responses API, not Chat Completions.
  • Rate limits start at 5,000 RPM on the Build tier; 429 errors require exponential backoff.

The Bottom Line

GPT-6.2 Astra is a pricing trap for the careless: the 20% sticker price cut is instantly erased by the 'max' reasoning tax and the 272K context cliff. Only migrate if you have the engineering discipline to monitor token usage and explicitly configure reasoning effort.