OpenAI has released GPT 6 Astra, its new flagship model, and published the full specification in its developer documentation on 3 September. The model takes a context window of 1,050,000 tokens, writes up to 128,000 tokens in a single response, and accepts both text and images. OpenAI describes it as its most capable model, built for the hardest end to end work.
The number that matters most to Australian buyers is not the context window. It is the price, which lands on exactly the same figure Anthropic charges for Claude Fable 5.1. That is not a rounding coincidence, and it says more about where the frontier market now sits than any benchmark chart.
What OpenAI actually shipped
Astra runs on the Responses and Chat Completions endpoints and supports batch processing, but it does not support fine tuning, embeddings, image generation or the realtime voice endpoints. Reasoning effort is adjustable across five levels, from low up to max, which lets teams trade cost against depth on each request. Supported features include streaming, structured outputs, function calling, file search, web search and prompt caching.
OpenAI points the model at complex reasoning, coding, computer use, research and document creation. Maximum input is 922,000 tokens, so the headline context figure includes the model’s own output budget rather than sitting on top of it. That distinction matters when you are sizing a job that pushes a large corpus through in one call.
The training data stops on 30 April 2026. Claude Fable 5.1 carries a June 2026 cutoff, so on recent events Anthropic is two months ahead. For most business workloads that gap is invisible. It matters for anything touching regulation, market data or product releases from the middle of this year, where the model will confidently describe a world that has already changed.

The price is not a coincidence
OpenAI charges $13.87 per million input tokens and $69.35 per million output tokens for Astra, with cached input at $1.39. Anthropic charges the same $13.87 and $69.35 for Claude Fable 5.1, with the same 128,000 token output ceiling and a context window of one million against Astra’s 1,050,000, according to Anthropic’s model documentation. Two rival flagships have converged on an identical price card.
There is a catch in the fine print. Prompts over 272,000 input tokens are billed at double the input rate and 1.5 times the output rate for the whole request, so a single long document push can cost several times what the headline figure suggests. Batch and flex processing run at half price, and fast mode at double.
OpenAI bills in United States dollars, so what an Australian team actually pays moves with the exchange rate. At the Reserve Bank rate on 4 September, with one Australian dollar buying 0.7210 United States dollars, the two models cost the same to run, which pushes the decision onto capability, availability and how each one behaves on your own evaluations rather than onto the invoice.
Who can use it today
Access opened faster than the launch suggested. On 3 September the documentation limited Astra to enterprises in OpenAI’s Trusted Access Program. By 4 September at 20:13 UTC that language had gone from OpenAI’s model page, which now mentions no Trusted Access, no waitlist and no staged rollout. Australian teams can call Astra today, on the Responses, Chat Completions and Batch endpoints.
The staged release lasted about a day. Anthropic ran the same pattern on 2 September, and we covered the Claude Fable 5.1 launch and who was let in first. Announcing a model and shipping it are still two separate events, but the gap here was measured in hours rather than weeks.
Trusted Access is OpenAI’s vetted enterprise channel, and it remains the first stop for a frontier release. What changed here is how briefly it held. A mid market Australian firm that read Thursday’s launch coverage and concluded it was locked out would have been wrong by Friday evening, which is its own argument for checking the model page rather than the announcement.
What Australian businesses should do now
The practical advice for an Australian business is still to move deliberately, but the reason has changed. Access is no longer the blocker, so evidence is. A specification sheet does not tell you whether Astra beats Claude Fable 5.1 on your documents, your tone and your edge cases, and nothing published so far does either.
What is worth doing now is building the comparison harness. Pick twenty real tasks from your own operations, run them against the model you use today, and keep the outputs. Then run the same twenty against Astra, which you can do today, and you will have a baseline instead of an opinion. Teams that skip this step tend to switch on vendor marketing and switch back three months later, and our breakdown of Claude and ChatGPT for business workloads is a reasonable starting shape for that harness.
For document heavy operations the one million token class of model genuinely changes what is feasible, because a full contract set or a year of correspondence now fits in a single request. That is also where the pricing trap above bites hardest, and where careful document automation design pays for itself rather than being an optional refinement.
A ceiling that keeps moving
Two flagship models, launched a day apart, at the same price, with the same output ceiling and context windows within five per cent of each other. The frontier is no longer a place where one vendor holds a durable lead for long, and the gap that used to justify a twelve month contract now closes in a fortnight.
For Australian buyers that convergence is good news, because it makes switching cheaper and makes any single vendor relationship less load bearing. The question worth asking is no longer which model is best, but how quickly your own stack can move when the answer changes. On current evidence, the answer changes about every eight weeks.
