GPT-6 Astra is live as of 3 September 2026. API model ID gpt-6-astra, $10 per million input tokens, $50 per million output, $1 per million cached input, batch at half price, and a Fast mode at 2× ($20/$100). The context window is 1,050,000 tokens with 922,000 max input and 128,000 max output. Two things are genuinely new: Astra operates software through screens rather than APIs (OSWorld 2.0: 72.6% vs 65.7% for GPT-5.6 Sol), and it is the first OpenAI model rated "Critical" on the cybersecurity axis of the Preparedness Framework, which means enterprise admins must switch it on manually.
OpenAI paired the launch with the phrase "the AGI era", which you can ignore. The pricing, the context arithmetic and the rollout gates are what change your week. Sources: OpenAI's launch materials as summarised by DataNorth and LLM-Stats, plus the model card figures quoted below.
Pricing next to the models you are already using
| Model | Input / MTok | Cached input | Output / MTok | Context |
|---|
| **GPT-6 Astra** | $10 | $1.00 | $50 | 1.05M |
| GPT-6 Astra Fast | $20 | — | $100 | 1.05M |
| Claude Fable 5.1 | $10 | $0.25 | $50 | 1M |
| Claude Opus 5 | $5 | $0.50 | $25 | 1M |
| Gemini 3.8 Flash | $0.75 (to 31 Dec 2026) | — | $3.75 | 1M |
Astra and Fable 5.1 have identical base prices. The gap is the cache: Anthropic cut Fable 5.1 cache reads to $0.25 on 1 September, four times cheaper than Astra's $1.00. For an agent loop that re-reads a 200k-token codebase on every turn, that is $0.05 per turn on Fable versus $0.20 on Astra before you generate a single output token. If your workload is cache-heavy, price it with the AI model cost calculator rather than the headline rate — the headline rate is the same and the bill is not.
Benchmarks: where Astra wins and where it does not
| Benchmark | GPT-6 Astra | Comparison |
|---|
| OSWorld 2.0 (desktop operation) | **72.6%** | GPT-5.6 Sol 65.7% |
| Terminal-Bench 4.0 | **57.9%** | Claude Fable 5.1 55.8% |
| DeepSWE v1.1 (bug fixing) | **74.1%** | Claude Opus 5 73.7% |
| Artificial Analysis Intelligence Index v4.1.1 | 61.2 | **Claude Fable 5.1 65.7** |
| ExploitBench | **100%** | GPT-5.6 Sol 78.5% |
| Time per agentic task | ~40 min | GPT-5.6 Sol ~75 min |
Read the pattern: Astra's wins are on tasks where the model drives a computer — desktop operation, terminals, repos — and by narrow margins on coding (0.4 points on DeepSWE, 2.1 on Terminal-Bench). On a broad reasoning index it trails Fable 5.1 by 4.5 points. The 47% reduction in time-per-task versus Sol is the number that will matter most in production: fewer steps means fewer tokens, which partly offsets the cache disadvantage.
There is a cost to fewer steps that OpenAI itself flags. Because Astra writes less reasoning per solution, its chain of thought is less monitorable than its predecessor's — OpenAI calls the decline "serious". If your safety story depends on reading the model's reasoning, that story got weaker with this release.
Screens, not APIs
The defining capability is computer use through the UI. Astra is built to see a screen, move a cursor and type, and it stayed within scope in 100% of OpenAI's honeypot tests, versus Sol straying in 48.2% of instances. For teams that have been wrapping every internal tool in a function schema so a model can call it, this inverts the integration cost: the model can use the tool the way a contractor would, through the interface that already exists.
Practical consequences. Legacy back-office software without an API becomes automatable. Your permission model has to move from "which functions can it call" to "which windows can it see and which accounts is it logged into" — a screen-driving agent with an admin session is an admin. And observability changes: you need screen recordings, not just tool-call logs. The Agent Ops bundle has the spec-and-observability templates we use for exactly this shift.
The "Critical" cyber rating and what it gates
Astra is the first model OpenAI has classified at the Critical level for cybersecurity: in evaluation it found and exploited previously unknown vulnerabilities in hardened systems from high-level instructions, without a human guiding each step, and discovered two zero-days along the way — including a browser compromise that escaped its sandbox and a privilege-escalation chain on a hardened OS. ExploitBench at 100% is the benchmark expression of the same thing.
The gates that follow from that rating:
Rollout is staged — a limited set of organisations on day one, then ChatGPT Plus, Pro, Business and Enterprise "over the coming days", then the API and AWS. Enterprise administrators must enable Astra manually; it is off by default. Deep cyber work goes through the Daybreak Blue access programme rather than the general API. Layered classifiers sit in front of the model, and OpenAI states plainly that the safeguards "may erroneously flag legitimate activity" — security tooling, pentest automation and CTF-style prompts should expect refusals and build a fallback route. Astra also declines 91.5% of jailbreak attempts, against 59% for Sol, which is a large jump in one generation.
We cover the industry-wide picture — Gemini 3.8 Flash Cyber and Claude Mythos 5.1 shipped the same week with their own gated tiers — in the cyber-capable model guide.
Migration notes for API code
Three things to change and one to test.
Context arithmetic. 922k max input is not 1.05M. If you were planning to stuff a million tokens of documents in, you cannot; leave room for the 128k output ceiling too. The token counter gives you a quick read on how much of that budget a document actually consumes under a GPT-family tokenizer.
Caching. At $1 per million cached tokens, prefix caching still pays back after one reuse of a large system prompt, so structure requests with the stable content first. But do the comparison honestly: on Fable 5.1 the same cached prefix is four times cheaper.
Fast mode. 2× price for up to 2× speed is a reasonable trade for interactive agents and a bad one for batch. Keep it off by default and enable per request.
Test the long-context notes feature. Astra ships an experimental mode that keeps searchable notes across context windows instead of a single rolling summary. It is the kind of thing that either quietly fixes your 3-hour agent runs or quietly changes what the model remembers. Run your longest evaluation with it on and off before trusting it.
Should you switch?
If your product is a computer-using agent — RPA replacement, QA automation, anything that drives a UI — Astra is the strongest option shipping today and the honeypot results are the reason. If your product is code generation inside an IDE or a CLI, the DeepSWE and Terminal-Bench margins are inside noise and the cache pricing decides it; Fable 5.1 is cheaper per turn for anything with a big stable prefix. If your product is high-volume chat, none of this applies — Gemini 3.8 Flash at $0.75 is a different price class entirely, and we compare it in the Gemini 3.8 Flash guide.
Quick answers
Can I call GPT-6 Astra from the API today?
The rollout is staged: a limited set of organisations on day one, then ChatGPT Plus, Pro, Business and Enterprise over the following days, then the OpenAI API and AWS. Enterprise administrators must switch the model on; it is off by default.
What is the GPT-6 Astra context window?
1,050,000 tokens in total, with a maximum of 922,000 input tokens and 128,000 output tokens per request. Budget for the input ceiling, not the total.
How much does GPT-6 Astra cost?
$10 per million input tokens, $50 per million output, $1 per million cached input. Batch processing is half price; Fast mode is double ($20 / $100) for up to twice the speed.
What does a Critical cybersecurity rating mean?
It is the highest tier on the cyber axis of OpenAI's Preparedness Framework: in evaluation the model found and exploited unknown vulnerabilities in hardened systems without step-by-step human guidance. It triggers extra classifiers, an access programme for deep cyber work and the admin opt-in.
Is Astra better than Claude Fable 5.1?
It depends on the task. Astra leads on Terminal-Bench 4.0 (57.9% vs 55.8%) and on desktop operation; Fable 5.1 leads on the Artificial Analysis index (65.7 vs 61.2) and its cache reads are four times cheaper. For cache-heavy agents the bill, not the benchmark, usually decides.
Whichever way you route, put the routing logic somewhere you can change it in a config file. The Claude Code Production Pack ships the model-routing rules we run across three providers, and the MCP server pack covers the tool layer that stays constant while models churn underneath. Every product mentioned is available at wowhow.cloud — pay once, ship forever.
Originally published at wowhow.cloud
Top comments (0)