GPT-6 Astra in 2026: OpenAI's agent-first flagship, priced per task
OpenAI's new flagship isn't a better chatbot — it's a computer-use engine priced at 2.5× the previous flagship. Whether that is a price hike or a discount depends entirely on how you buy it.
- Flagship rate — GPT-6 Astra lists at $10 input / $50 output per million tokens — 2.5× GPT-5.6 Sol's promotional rate, matching Anthropic's Fable 5.1.
- Built for agents — Positioned for computer use and software engineering: 1.05M context, 128K max output, and top scores on OSWorld and Terminal-Bench 4.0.
- Tops the coding board — Ranked #1 for coding on the LLM Stats leaderboard as of September 12, 2026, with 74.1% on DeepSWE 1.1.
On September 3, 2026, OpenAI shipped GPT-6 Astra (API model ID gpt-6-astra) — four weeks after saying it had slowed the model over cyber risk. The positioning is explicit: computer use and software engineering, not chat. OpenAI calls it "a new frontier in the speed, accuracy, and safety of computer use" and its best model for software engineering to date.
Specs, as shipped
| Spec | GPT-6 Astra |
|---|---|
| Context window | 1,050,000 tokens |
| Max output | 128,000 tokens |
| Input / output | Text + images in, text out (no audio/video at launch) |
| Knowledge cutoff | April 30, 2026 |
| Reasoning effort | Low, medium, high, xhigh, max |
| Tools | Web search, file search, code interpreter, hosted shell, computer use, MCP |
Specs per launch coverage, September 2026. Rollout is staged: trusted-access organizations first, then ChatGPT tiers, the API, and AWS Bedrock.
The price card — and the four prices inside it
The headline is $10 per million input tokens and $50 per million output, with cached input at $1. But Astra really has four prices:
- Standard: $10 / $50.
- Batch / Flex: half price — the patient buyer's rate.
- Fast mode: double price for up to double speed.
- Long context: any prompt past 272K input tokens reprices the entire request to $20 / $75.
OpenAI's own framing, voiced by president Greg Brockman at launch: stop counting tokens and judge price per completed task. The numbers give that argument legs — Astra reportedly finishes computer-use tasks faster than Sol, so fewer, quicker runs can undercut a cheaper per-token model on total bill.
Benchmarks, with the required caveat
Every figure below is OpenAI's own, published at launch, with no independent replication yet. Treat them as a vendor claim to verify on your workload:
- Terminal-Bench 4.0: 57.9% (Sol 37.3%) — the long-horizon agent test.
- DeepSWE v1.1: 74.1% (Sol 72.7%).
- OSWorld 2.0: 72.6% (Sol 65.7%) — autonomous computer use.
- FrontierMath Tier 4: 97.6% (Sol 83.0%).
- Hallucination rate: 4.2% vs Sol's 12.2%.
Independently, the LLM Stats coding leaderboard (updated September 12, 2026) ranks GPT-6 Astra #1 with a 48.5 rating, logging 74.1% on DeepSWE 1.1 and 53.3% on FrontierCode 1.1.
The cyber footnote that delayed it
Astra is the first model OpenAI rates Critical for cyber risk under its Preparedness Framework — the reason for the four-week delay. Cyber-sensitive capabilities sit behind Daybreak, a gated trusted-access program; the public model refuses advanced cyber tasks. Eligible API customers can run it with Zero Data Retention. If you are in a regulated industry, this paragraph matters more than any benchmark.
Stay on GPT-5.6 Sol at $4/$20 while promotional pricing holds (guaranteed through at least November 21, 2026) for everything else — and run your own head-to-head before that deadline decides for you.
How to decide before November 21
Run the same three agent tasks on Astra and Sol, log tokens and wall-clock time, then price both at standard versus Batch rates. If Astra's speed advantage doesn't erase its 2.5× premium on your tasks, the flagship rate is a tax on impatience.
Prices and benchmark figures are standard published rates as of September 12, 2026 and can change without notice. Some outbound links may be affiliate links — see our disclosure. Found an error? Tell us and we'll correct it in public.