OpenAI's developer documentation for GPT-6 Astra carries a pricing rule that most launch coverage skipped, and it inverts the economics of exactly the workloads the model is sold for. The advertised $10 per million input tokens and $50 per million output holds only below a threshold, and past it the whole request reprices.
The threshold applies to everything, not the excess
The model page states it plainly: "Prompts with more than 272K input tokens are priced at 2x input and cache rates and 1.5x output for the full request." The words for the full request carry the cost. A 273,000-token prompt is not billed at the standard rate with a surcharge on the last thousand tokens — all of it reprices, to $20 input and $75 output. A prompt two thousand tokens shorter costs half as much.
Cache writes cost more than fresh input
The same table prices cached input at $1.00 per million and cache writes at $12.50 — the page notes writes bill at 1.25x the uncached input rate. Prompt caching is normally presented as a saving; here it is an upfront premium that only pays back on reuse. Batch and Flex run at half the standard rate; Fast mode runs at double, and the page adds that Fast mode "is unavailable for GPT-6 Astra with EU data residency."
Controls the model no longer accepts
OpenAI's changelog lists the limitations: no support for the none reasoning-effort level, no custom temperature or top_p, and no logprobs. That is not a rounding-off of rarely used parameters. Evaluation harnesses, classifier pipelines and anything doing token-probability work cannot port to Astra at all, whatever its benchmark scores say.
What the received framing gets wrong
Trackers and write-ups print $10/$50 flat on a million-token model. Both halves mislead. The price applies to the short-context case, while agentic and computer-use runs — the ones OpenAI markets Astra for — live above the threshold and are a $20/$75 model. And the "1M context" figure is the 1,050,000-token context window stated on the page, which includes generation headroom; with 128,000 max output tokens also stated there, usable input works out around 922,000. OpenAI does not publish that subtraction, and anyone sizing a document pipeline off "1M input" is over by roughly 128K.