OpenAI's newest model has a 1,050,000 token context window with a pricing cliff two thirds of the way in, and a reasoning effort scale that gained two levels at the top and lost one at the bottom. Both defaults will quietly cost you money if nobody changes them.
GPT-6 Astra shipped this week with a documented context window of 1,050,000 tokens and a 128,000 token maximum output, and the headline number is the part most teams will get wrong. OpenAI's API documentation states that prompts exceeding 272,000 tokens are charged at twice the input and cache rates and 1.5 times the output rate. The window is not a flat rate resource. It has a step in it at roughly 26 percent of its own advertised size, and nothing in the API tells you when you crossed it.
Price the boundary before you design across it. At the standard rate of 10 dollars per million input tokens, a 250,000 token prompt sits under the threshold and costs about 2.50 dollars to send. A 500,000 token prompt crosses it, and OpenAI's documentation is explicit that prompts above 272,000 tokens are charged at 2x input and cache rates, which puts that request somewhere between roughly 7 and 10 dollars depending on how the tier is applied. Either way the prompt doubled and the bill did considerably more than double. Run that inside an agent loop that resends a growing transcript on every turn and the curve gets ugly quickly. If your workload genuinely needs to live above the threshold, confirm with OpenAI whether the multiplier applies to the whole prompt or only the excess, because the answer changes your forecast by thousands of dollars a month at volume.
The cache is where the money actually is, and Astra's cache pricing rewards discipline. Cached input costs one dollar per million tokens against 10 dollars for uncached, a tenfold difference, while cache writes cost 12.50 dollars per million. That spread means a prefix you reuse pays for its write on the second read and prints money after that, but a prefix you write and never reuse costs you 25 percent more than simply sending it. Cache the stable parts of your prompt, the system instructions, the schemas, the reference documents, and keep the volatile parts after them so a small change does not invalidate the whole prefix.
The reasoning effort scale moved in both directions. OpenAI's documentation lists xhigh and max above high, which is two more levels of deliberation than the previous generation offered, and the model guidance states that GPT-6 Astra does not support the none effort at all. If your code carries a hardcoded reasoning effort of none or minimal from a GPT-5 era integration, it will not do what it did before. OpenAI's own migration guidance is to start at low and compare results rather than assuming the nearest label preserves your behavior. The two new levels at the top are worth deliberate experiments, not a blanket upgrade: max effort spends more output tokens, and output tokens are the expensive side of the meter at 50 dollars per million.
One last lever most teams forget. Batch and Flex service tiers run at 50 percent of standard rates, and fast mode costs double. A nightly evaluation suite, a bulk document classification job, a backfill: none of those need interactive latency, and all of them are currently paying interactive prices at most companies. Moving that traffic to Batch is the single largest cost reduction available on this model that requires no change to a prompt.
Try it today
- Instrument your token counts before you migrate, and alert on any single request crossing 272,000 input tokens so the 2x billing tier never surprises a monthly invoice.
- Restructure prompts so stable content comes first and volatile content last, then verify cache hit rates, because cached input at one dollar per million against 10 dollars uncached is the largest lever on this model.
- Audit every hardcoded reasoning effort value in your codebase for none and minimal, which GPT-6 Astra does not accept, and start at low rather than guessing the equivalent.
- Test xhigh and max on a narrow set of genuinely hard tasks and measure output token consumption, since output bills at 50 dollars per million and higher effort spends more of it.
- Move every non interactive workload, evaluation runs, batch classification, backfills, to the Batch or Flex tier at 50 percent of standard rates before optimizing anything else.
- Confirm the April 30, 2026 knowledge cutoff against your retrieval design, and make sure anything more recent than that date reaches the model through context rather than assumption.