OpenAI released GPT-6.2 Astra and its cheaper sibling, GPT-6.2 Sol, on November 17, 2026, marking a significant shift in how developers interact with the company's reasoning models. While the model ID changed from gpt-6-astra to gpt-6.2-astra, the most critical change for existing users is that the API now strictly rejects "none" as a reasoning effort value. This means every single call incurs reasoning token costs, fundamentally altering the economics of lightweight requests that previously bypassed the reasoning chain.
The New Reasoning Effort Landscape
The Responses API now requires explicit effort settings, with verified support for low, medium, high, xhigh, and max. OpenAI has not specified a default effort level, forcing developers to make a conscious choice for every request to ensure reproducibility and cost control. Simon Willisonβs launch-day benchmarks highlight the volatility: on Astra, effort "high" cost 11.55 cents, "xhigh" cost 17.70 cents, and "max" spiked to 88.71 cents for the same image prompt. Moving from "high" to "max" nearly octupled the cost, demonstrating that higher effort levels in GPT-6.2 are significantly more expensive than in previous generations, even if per-token pricing dropped by 20%.
Pricing Cliff at 272K Tokens
Developers must be wary of a harsh pricing mechanism tied to context length. If a request exceeds 272,000 input tokens, the pricing for the entire request changes: input and cache prices double, and output prices increase by 1.5x. For example, a 300K token prompt is billed at $16 per million input tokens and $60 per million output tokens, rather than the standard $8 and $40. This applies to the whole request, not just the tokens over the threshold, making large repository uploads or long-context agent loops potentially prohibitively expensive if not carefully chunked or cached.
Mixed Reception and Migration Risks
Early reviews of GPT-6.2 Astra are polarized. Ethan Mollick described the model as very inconsistent and weird, while Dan Shipper noted frequent early stopping in end-to-end workflows. Conversely, other testers praised its performance in code review and computer use tasks. Migration from GPT-6 Astra is mostly a one-line code change, but developers must audit their codebases to ensure no paths send "effort": "none," which now results in an error. Additionally, both new models have reached the "Critical" cybersecurity threshold, meaning stricter safety filters may cause previously successful prompts to be rejected.
Key Takeaways
- The model ID is gpt-6.2-astra; the cheaper variant is gpt-6.2-sol.
- Reasoning effort "none" is no longer supported; all calls consume reasoning tokens.
- Input costs double and output costs increase by 1.5x for requests exceeding 272K tokens.
- Batch API offers a discount for non-urgent evaluations and backfills.
- Context window is 1,050,000 tokens total, with a max input of 922,000 and output of 128,000.
The Bottom Line
GPT-6.2 Astra is a powerful but treacherous tool for the unprepared. The removal of the "none" effort option and the 272K pricing cliff mean that naive migration from GPT-6 Astra could result in surprise bills. Developers must implement strict token counting and effort controls before pushing this model to production.