GPT-6 Astra: a million-token context with a long-request surcharge

OpenAI released GPT-6 Astra on September 3. It has a context window of 1,050,000 tokens. Context is the total information in one request, including space for the response; a token roughly corresponds to a word fragment. The exact count depends on the text and language.

That capacity allows a contract to be sent with appendices, correspondence and earlier versions. Astra’s maximum request size and pricing threshold differ, however: the surcharge starts well before the context fills up.

OpenAI’s model page lists base rates of $10 per million input tokens and $50 per million output tokens. Above 272,000 input tokens, input and cache rates double and output rates increase by half. The surcharge covers the entire request, not just the text above the threshold.

For example, 500,000 uncached input tokens cost $10, and 20,000 output tokens cost another $1.50: $11.50 at the standard rate before additional tools. This prices a specified volume; it does not establish how many tokens a particular task will consume.

An original contract and an amendment can specify different payment deadlines. If retrieval sends only the original clause to the model, the change is absent from its request. Sending the related documents in full allows the model to see both versions.

This explains the usefulness of a large context for document comparison. A question about one contract does not require every customer’s archive, however: the contract and its related amendments can be enough. Retrieval needs to preserve those connections, or it may discard the relevant change along with unnecessary text.

OpenAI lists separate cache rates for stored text that can be reused. A base-rate read costs $1 per million tokens, rising to $2 when the request exceeds 272,000 input tokens. Writes cost $12.50 and $25 per million respectively.

Caching is particularly relevant when working with one large document set: the text stays unchanged while questions vary. The reused portion of input is billed at the read rate, while new text uses the ordinary input rate. Each response is generated and billed again.

For the same 500,000 tokens, reading fully cached input costs $1 rather than $10 for new input. The 20,000 output tokens still cost $1.50. This comparison excludes the initial cache write, new question text and tools. The read discount does not remove the long-request surcharge: the size still exceeds 272,000 tokens.