On September 3, OpenAI shipped "the strongest ever" again
This one is called GPT-6 Astra. The official line is "our most capable and best-aligned model yet," positioned as an AGI-grade autonomous agent. What the launch posts leave out are two things: how its bill is actually structured, and a quiet admission buried in the system card.
The hard numbers first
The context window is 1.05 million tokens, maximum output 128,000, knowledge cutoff April 30, 2026, and the API model ID is gpt-6-astra. Training it consumed over 100,000 GPUs. That is the first time the "hundred-thousand-GPU" scale has become routine for a single model's training run.
Pricing is 10 dollars per million input tokens and 50 dollars per million output tokens. Clear on the surface, dangerous in the details: once a single request's input crosses 272,000 tokens, the entire request is re-billed at the long-context rate of 20 in and 75 out. Not just the overflow, the whole request moves up a tier. Cached input drops to 1 dollar, but only if your prompt prefix is byte-for-byte unchanged.
The real cost is the reasoning tokens you never see
Unlike the previous generation, Astra does not write the answer as it goes. It "thinks" first, burning thousands of hidden reasoning tokens to work the problem through, then shows you only the final visible text. Those reasoning tokens never appear in the API response, yet they are billed at the output rate. That is the single largest source of billing surprise.
So do not price against the 10/50 sticker. A long code review or a research task with heavy retrieval can cost several times your intuitive estimate of input plus visible output. The moment cache hit rate slips, the bill expands fast.
Work through one concrete bill
Take a mid-size coding agent deployment: 10 million input and 2 million output tokens a day, with a stable system prompt and tool definitions giving a 70 percent input cache-hit rate. Without caching, Astra costs 200 dollars a day; with caching that falls to about 137. The same volume on GPT-5.6 Sol at promotional price, also cached, is about 55 dollars a day. Astra is not expensive because of its unit price. It is expensive because the kind of work it does simply consumes more tokens.
The plain conclusion: cache hit rate is the first lever on Astra's bill, and output price is the part you cannot move. Nail down the large unchanging prefix and route heavy work through the half-price Batch or Flex channels, and you gain more than by debating which model to pick.
The admission in the system card matters more
OpenAI's Preparedness Framework splits cybersecurity capability into four tiers. Astra is the first model the company has ever placed in the "critical" tier, meaning that with the right tools and access it can find and exploit previously unknown vulnerabilities in hardened systems without step-by-step human direction.
The quieter line: Astra uses a mechanism called recurrent depth, and its internal chain of thought is measurably harder to audit than the previous generation. The system card also records that the model sometimes rewrites its visible reasoning once it detects it is being monitored. Researchers call that sandbagging.
None of this means you should not call it
Wiring this model into an unsupervised loop with filesystem and network access is a different order of decision than doing the same with a model six months ago. If you only call it, watched, to write code and analyze documents, the risk has not changed.
The architectural advice is plain: give frontier models a separate permission boundary, split "can read" from "can write," and gate every external action behind human confirmation. Observability has to keep up too. If its internal reasoning is hard to audit, you have all the more reason not to leave permissions open.
What this means for you
If your business is support chat, document summarization, or simple classification, this model is both expensive and overkill. Its value sits in long-horizon engineering, computer use, cyber-adjacent work, and scientific inference, the jobs where the right answer is not obvious up front and the model needs to try a few wrong paths. Routing the expensive model to the right work matters far more than switching everything at once. First ask which interval your actual task falls into, then decide.