OpenAI has quietly changed how Codex bills for long context, and GPT-6 Astra is the exception. Its native window stretches to about 1.05 million tokens, and in Codex, inputs that blow past 272K no longer set off long-context pricing. You can approach a million tokens of context and still pay standard rates.
Other models don't get that break. GPT-5.4, GPT-5.5, and GPT-5.6 still trigger the old multiplier when context goes above 272K: 2x for input and cached input, 1.5x for output.
There is one important limitation: this special pricing only applies inside Codex. Via the API, OpenAI is still charging 2x input and 1.5x output once you cross the long-context threshold.