TL;DR
On September 3, 2026, OpenAI released GPT-6 Astra — its most capable and aligned model yet, trained on 100K+ GPUs at the Stargate site in Texas. It is OpenAI’s first model formally rated “Critical” for cyber capability under its Preparedness Framework.
The headline numbers: $10 input / $50 output per million tokens, a ~1.05M-token context window, 99.9% on ARC-AGI-3, and an ExploitBench score of 100% (in its restricted Daybreak form). It is the model behind Greg Brockman’s “Welcome to the AGI era.”
But the rollout is phased: Daybreak partners and a first wave of enterprises get it first, paid ChatGPT plans and the public API follow within days, and the free tier has no date. Enterprise workspaces have it disabled by default.
For developers the honest takeaway: Astra is a real leap on long-horizon agent work and science/terminal tasks, but its gains are not universal (it trails Claude Fable 5.1 on general tool reasoning), the pricing has a 272K-token trap, and every number here is vendor-reported.
What is GPT-6 Astra?
GPT-6 Astra is OpenAI’s next-generation flagship, succeeding the GPT-5.6 family (Sol/Terra/Luna). It was trained with over 100,000 GPUs at the Stargate supercomputing site — OpenAI’s first flagship to use previous-generation models extensively in supervised training.
Key specs from the launch:
| Spec | GPT-6 Astra |
|---|---|
| Model ID | gpt-6-astra |
| Context window | ~1.05M tokens (92.2K max input / 12.8K max output shown; long-context variants) |
| API price | $10 in / $50 out per M |
| Reasoning effort | low → medium → high → xhigh → max |
| Interfaces | Responses API, Chat Completions, Batch; AWS Bedrock later |
| Not supported | Realtime API, Assistants API, fine-tuning, embeddings, native image/video gen |
| Cyber rating | Critical (first for OpenAI) |
Can you use it today?
Not by default. The rollout is deliberately phased:
- Day 1 (Sep 3): Daybreak Blue partners (Cisco, Cloudflare, Palo Alto Networks) and the first wave of approved enterprise customers.
- Next few days: Paid ChatGPT plans (Plus, Pro, Business, Enterprise) and the public
gpt-6-astraAPI, then Amazon Bedrock. - Free tier: No announced date.
Enterprise rule: Astra is disabled by default in workspaces — an admin must explicitly enable it after reviewing the system card. If your org uses OpenAI, check whether your workspace has been added to the rollout list before assuming you can call it.
Pricing & the 272K trap
Standard pricing per million tokens:
| Item | Price / 1M tokens |
|---|---|
| Input | $10.00 |
| Output | $50.00 |
| Cached input | $1.00 |
| Cache write | $12.50 |
| Batch / Flex | ~50% |
| Fast mode | 2.5x speed at ~2x price ($20 / $100) |
The 272K-token trap: when a single request exceeds 272K input tokens, input and cache prices double, and output rises 50% — and the penalty applies to the entire request, not just the excess. A long-context session that crosses this threshold can cost dramatically more than expected. Plan to keep single inputs under 272K or use Batch for huge jobs.
Benchmarks: honest picture
Astra’s scores are mostly at its highest reasoning settings (xhigh/max) — not equal-cost comparisons:
| Benchmark | GPT-6 Astra | GPT-5.6 Sol |
|---|---|---|
| ARC-AGI-3 | 99.9% | 7.8% |
| FrontierMath (Tier 4) | 97.6% | 83.0% |
| ExploitBench | 100% (Daybreak form) | 78.5% |
| Terminal-Bench 4.0 | 57.9% | 37.3% |
| OSWorld 2.0 | 72.6% | 65.7% |
| Coding-Agent Index v1.4 | 67.0 | 65.1 |
| HLE (with tools) | 57.2% | 65% (Sol higher) |
The honest caveats: (1) Astra is not universally better — on Humanity’s Last Exam with tools it trails GPT-5.6 Sol, and on general tool reasoning OpenAI itself acknowledges Claude Fable 5.1 leads (65.0% vs 57.2%). (2) The Artificial Analysis Coding Index shows Astra at 67, slightly below Claude Fable 5.1’s 68.1. (3) Everything here is vendor-reported and unaudited.
vs GPT-5.6 Sol: should you switch?
The practical question. GPT-5.6 Sol is still on its promo price of $4/$20 (until Nov 21, 2026), making it ~2.5x cheaper than Astra’s $10/$50.
Route to Astra if you do long-horizon agent work, science/terminal tasks, complex persistent refactors, or need the cross-context memory that Astra enables in Codex. Its Terminal-Bench 4.0 (57.9% vs 37.3%), SRE-Bench (88% vs 55.9%) and OSWorld gains are genuinely large.
Stay on Sol if you want predictable low cost for high-volume work. For cost-sensitive production routing, many teams should run Sol for routine tasks and Astra only for the hardest subset. A 2.5x price premium is only worth it where Astra’s accuracy actually wins your workload.
If you already use Claude Opus 5 or Fable 5.1: don’t migrate on day one. Run a private A/B on your own test set first.
How to use it
When your account has access, the API is straightforward. Tool/function calling requires the Responses API.
from openai import OpenAI
client = OpenAI()
response = client.responses.create(
model="gpt-6-astra",
reasoning_effort="medium", # start low, not "max"
input="Analyze this Python code for concurrency bugs and fix them.",
tools=tools_list, # must use Responses API
)
# Escalate reasoning mid-session, reusing the prefix:
response = client.responses.create(
model="gpt-6-astra",
previous_response_id=response.id,
reasoning_effort="xhigh",
)Practical tips from launch analysis: start at medium reasoning, keep single inputs under 272K tokens to avoid the whole-request price jump, and use Batch for large jobs. In Codex, Astra can now “take notes across context windows” instead of compressing long context — so older conversation details stay retrievable on long-horizon tasks. If you built prompt-chunking workarounds to avoid summary loss, verify whether they’re still needed.
Cyber & guardrails
This is the unusual part of the Astra launch. It is OpenAI’s first model rated Critical for cyber capability — able to autonomously discover unknown vulnerabilities and chain exploits. OpenAI has acknowledged that monitoring Astra’s reasoning is harder than for any previous model.
The rollout therefore ships with unusual constraints: phased access, enterprise-disabled-by-default, Daybreak Blue/Red separation, a chain-of-thought classifier, and runtime interruption layers that can surface anomalous tool calls for human confirmation. The public model is more restricted than the Daybreak form that hit ExploitBench 100%.
For developers running agents: treat Astra like any powerful agent — sandbox code execution, apply least-privilege, require human approval for production or external-network writes, and keep audit logs. Don’t treat it as a fully trusted black box.
How we wrote this
This piece was compiled on September 4, 2026, the day after GPT-6 Astra’s release. We cross-referenced:
- OpenAI’s launch materials and system card (via coverage in the Verge, TechCrunch, Bloomberg, and Wired).
- Independent technical breakdowns (fanweibin.cn, syntaxandsignal.tech, aipricing.guru) for API pricing, the 272K-token rule, and rollout sequencing.
- Comparison data for GPT-5.6 Sol, Claude Fable 5.1, and Gemini 3.8 Flash from this week’s W37 news issue and vendor-published benchmarks.
We have not independently tested GPT-6 Astra — it was still phasing out at the time of writing. All benchmark numbers are vendor-reported and unaudited; rollout timing and pricing may change. Verify access and current rates in your OpenAI account before budgeting.
Corrections or counter-evidence welcome via our about page.