OpenAI's new generation did not add a single token of context. It costs two and a half times more

GPT-6 Astra, released on 3 September 2026, is the first model of OpenAI's sixth generation and, by the company's own description, its most capable. Put its model card next to the one for GPT-5.6 Sol, the flagship it succeeds, and the headline numbers are the same numbers. Both models have a 1,050,000-token context window. Both cap output at 128,000 tokens per request. Both price prompts above 272,000 input tokens at 2x input and 1.5x output for the whole request - Astra additionally doubles the cache rate at that threshold. The one specification that genuinely moved is the knowledge cutoff, from 16 February 2026 to 30 April 2026: ten and a half weeks of fresher training data. The price moved further. Sol runs at $4 per million input tokens and $20 per million output; Astra at $10 and $50. That is 2.5 times the cost per token in both directions. Sol's rate is promotional - OpenAI guarantees it at least through 21 November 2026 - but it is the rate a developer pays today, and it was itself a cut of 20% on input and 33% on output announced in August. Astra also takes options away. It rejects custom temperature, top_p and log probabilities, all of which Sol accepts. It drops the 'none' reasoning effort level, so the model always reasons. Tool calling works only through the Responses API; calling tools from Chat Completions requires migrating. In exchange it adds five effort levels up to xhigh and max, effort changes mid-conversation, asynchronous tool calling, mid-turn steering over WebSockets, and asynchronous misalignment monitoring that can stop an agent run for review. OpenAI has published no benchmark figures alongside the model card, so the case for paying 2.5 times more currently rests on the company's own description of the model as its most capable, and on capabilities aimed squarely at long agent runs rather than at bigger prompts.