OpenAI's GPT-6 Astra is now generally available outside OpenAI's own products. GitHub made it GA in Copilot on September 4, 2026, and AWS made it GA on Amazon Bedrock on September 8, 2026. On Bedrock it accepts a context window of up to 1 million input tokens.
For developers the number that matters is the bill, and OpenAI's pricing is more complicated than a single rate.
What it costs
OpenAI's pricing documentation splits GPT-6 Astra two ways at once. Prices change with the length of your context, and again with whether you pick Standard or Fast mode. All figures are per million tokens.
| Mode and context | Input | Cached input | Output |
|---|---|---|---|
| Standard, short context | $10.00 | $1.00 | $50.00 |
| Standard, long context | $20.00 | $2.00 | $75.00 |
| Fast mode, short context | $20.00 | — | $100.00 |
| Fast mode, long context | $40.00 | — | $150.00 |
Two details are easy to miss and expensive to discover late.
Fast mode doubles the input price and doubles the output price against Standard. It is also not available everywhere: OpenAI's documentation states that "Fast mode is unavailable for GPT-6 Astra with EU data residency. Use Standard processing for those requests."
Separately, models released after March 5, 2026 carry a 10% cost increase when called through regional endpoints. GPT-6 Astra is inside that window, so a regional deployment costs more than the table above.
Cached input is where the real savings sit. At $1.00 against $10.00, a cache hit on short context costs a tenth of a fresh read. That pattern is now standard across frontier providers, and Anthropic moved the same lever when Claude Fable 5.1 cut cache reads by 75% while holding token prices flat.
Where you can get it
| Platform | Details |
|---|---|
| Amazon Bedrock | GA September 8, 2026. Up to 1M input tokens. ChatGPT Work and Codex can be configured to use it here |
| GitHub Copilot | GA September 4, 2026, via the model picker |
| Copilot plans | Pro+, Max, Business and Enterprise |
| Copilot surfaces | VS Code, Visual Studio, Copilot CLI, the coding agent, the Copilot app, github.com, GitHub Mobile, JetBrains IDEs, Xcode and Eclipse |
GitHub describes the Copilot rollout as gradual, so the model may not appear in your picker on day one. Business and Enterprise administrators control access through model policies, which means an individual developer on a company plan may be waiting on an admin rather than on GitHub.
GitHub also states that Copilot use is billed as "usage-based billing at OpenAI's provider list pricing." The rates in the table above are the rates you pay there, not a separate Copilot tariff.
AWS points to the Bedrock documentation for region availability rather than listing regions in the announcement, and neither announcement states rate limits.
What this means for developers
Price your own workload before switching anything. The four-way split between context length and mode means a single headline rate tells you very little. On output, a long-context Fast mode call costs three times the short-context Standard rate.
Check whether you are actually short context. OpenAI's documentation prices short and long context differently but the announcements do not state where the boundary sits. If your prompts hover near a threshold you cannot see, your per-call cost can change without your code changing. That is worth measuring against real traffic before you commit a budget.
If you are in the EU, plan around Fast mode rather than for it. EU data residency rules it out entirely, and a regional endpoint adds 10% on top. An architecture that assumes Fast mode will not survive a compliance review that requires data residency.
The caching advice is the least glamorous and the most valuable. Structure prompts so the stable part comes first and stays byte-identical between calls, because that is what makes a cache hit possible. At a tenth of the input price, getting this right is worth more than most model choices.
One caveat on capability. Astra's published benchmark results have varied sharply with how it is tested, including a 37-point gap on ARC-AGI-3 that came down to the test harness. General availability changes what you can buy, not what it can do. Run your own evaluation on your own tasks before moving production traffic.
This article was first published on Tech AI Wire.
Related on Tech AI Wire
- Claude Fable 5.1 cuts cache reads 75% and keeps token prices flat
- GPT-6 Astra scores 99.9% or 62.7% on ARC-AGI-3, depending how you ask
Sources
- OpenAI GPT-6 Astra is now generally available on Amazon Bedrock - AWS
- GPT-6 Astra is generally available in GitHub Copilot - GitHub Changelog
- OpenAI API pricing - OpenAI
Top comments (0)