WHAT DEVELOPERS ACTUALLY GET
● Model string: gpt-6-astra. Also on Amazon Bedrock and Microsoft Foundry.
● Fast mode: twice the standard speed at twice the standard price.
● Three new primitives: asynchronous function calls, mid-turn steering, and reasoning-effort changes that do not invalidate the prompt cache.
The third one is the one that matters
Prompt caching is how you make repeated calls against a large fixed context affordable. Cached tokens cost a fraction of fresh ones, so anything that invalidates the cache resets you to full price.
Historically, changing how hard a model thinks meant changing the request in a way that broke the cache. So you either committed to one reasoning level for the whole session, or you paid full rate every time you changed your mind.
WHAT THIS UNLOCKS
You can now run cheap shallow reasoning across most of a task and turn it up only for the steps that need it, without losing the cache on your context.
For a long agent session against a large codebase, that is a structural cost change rather than a marginal one.
Mid-turn steering
The ability to redirect a model partway through a response, rather than waiting for it to finish and starting again.
For interactive work that is a latency improvement. For agent workflows it is more than that — an agent heading down the wrong path can be corrected at step three rather than at step twenty, which is where most wasted spend in long sessions actually goes.
Asynchronous function calls
Tool calls that do not block. The model can issue a request and continue rather than waiting for the result.
This matters most where tools are slow — a database query, an external API, a build. Sequential blocking calls are why some agent tasks take minutes that look like they should take seconds.
The pricing detail to verify
| What is known |
Detail |
| Model string |
gpt-6-astra |
| Fast mode |
2x speed, 2x price |
| Also available on |
Amazon Bedrock, Microsoft Foundry |
| Standard per-million rate |
Check the pricing page directly. It was not published at launch and figures in circulation vary |
Worth being explicit about that last row. Fast mode being double the standard price tells you nothing useful until you know the standard price, and third-party listings appearing this week are not consistent with each other. Verify at source before budgeting.
What to do this week
- Do not migrate yet. Wait for confirmed pricing and at least one independent benchmark. Both usually arrive within a fortnight of a launch.
- If you use prompt caching heavily, the reasoning-effort primitive is the thing to test first. It changes the economics of long sessions more than the model does.
- If you run agents with slow tools, async function calls are worth prototyping against your actual bottleneck.
- Compare Fast mode honestly. Double price for double speed is only worth it where latency has a cost you can name.
Sources
FAQ
What is the Astra model string?
gpt-6-astra. It is also available through Amazon Bedrock and Microsoft Foundry.
How much does it cost?
Per-million pricing was not published at launch and third-party figures are inconsistent. Fast mode runs at twice the standard speed for twice the standard price. Check the OpenAI pricing page directly.
What is mid-turn steering?
The ability to redirect the model partway through a response rather than waiting for it to finish. For agents, it means correcting a wrong path at step three instead of step twenty.
Why does the prompt cache detail matter?
Changing reasoning effort previously invalidated the cache, so you either fixed one level for a session or paid full rate to change it. Now you can vary effort per step while keeping cached context cheap.
What are asynchronous function calls?
Tool calls that do not block, so the model can continue while waiting on a slow tool. Most useful where tools are genuinely slow, such as database queries or builds.
Should I migrate now?
Not yet. Wait for confirmed pricing and independent benchmarks, both of which typically appear within two weeks of a launch.