Updated September 26, 2026 · 7 min read
GPT-6 Astra API: the developer's guide
Astra's API is powerful and expensive. Model ID gpt-6-astra, roughly a million tokens of context, $10/$50 per million tokens — and a handful of new capabilities that change how you architect agents. Here's what matters.
Model specifications
| Parameter | Value |
|---|---|
| Model ID | gpt-6-astra |
| Context window | 1,050,000 tokens |
| Maximum output | 128,000 tokens |
| Inputs | Text, images |
| Reasoning effort | low, medium, high, xhigh, max |
| Streaming / Function calling / Structured Outputs | Supported |
| Fine-tuning | Not supported |
| Weights | Closed |
Pricing
| Tier | Input / 1M | Output / 1M |
|---|---|---|
| Standard | $10 | $50 |
| Cached input | $1 | — |
| Batch | $5 | $25 |
| Fast mode | $20 | $100 |
New capabilities worth knowing
- Asynchronous tool calls. The model keeps reasoning while your app executes a tool — no more idle waiting during side effects.
- Mid-turn steering. Over a WebSocket connection, users can change requirements mid-execution; completed work is preserved instead of thrown away.
- Hot-swappable reasoning effort.
configuration_updatechanges effort level within an ongoing conversation without resetting the cache — expensive steps get "max", trivial ones get "low". - Long-term memory in Codex. An experimental notes-and-retrieval mechanism across context windows reduces information loss from repeated compaction.
- Production monitoring. Tool-use trajectories are checked asynchronously for potentially unauthorized actions.
Cost-saving patterns that actually work
- Model routing. Reserve
gpt-6-astrafor genuinely hard steps; route everything else to cheaper models. Teams report this cuts Astra spend dramatically — one partner saw up to 20% fewer tokens versus other frontier models on the same workflows. - Cache aggressively. Cached input at $1/1M makes repeated system prompts and large reference contexts nearly free. Structure prompts so the stable part comes first.
- Batch non-urgent work. Half price ($5/$25) for anything that can wait.
- Right-size reasoning effort. The gap between "low" and "max" is the biggest cost lever in the API. Most production calls don't need "max".
Note: Astra's API costs roughly 2.5× GPT-5.6 Sol on both input and output. It earned that premium on computer-use and agentic benchmarks — but applying it to high-volume, low-complexity traffic is the fastest way to a shocking invoice. Full numbers in our pricing guide.
Availability
Astra is on the OpenAI API and rolling out to AWS, and it's integrated into partner harnesses (Cognition's Devin on launch day, Cursor, and others). Check your cloud provider's model catalog for regional availability.