Last updated: September 4, 2026 · By Vishal Swami, Founder & Lead AI Reviewer, AISagely
OpenAI begins rolling out GPT-6 Astra on September 3, 2026, starting with a small group of enterprise cybersecurity partners before it reaches ChatGPT and the API over the following days. It's not available to everyone yet, and the plan you're on decides whether you can even see it this week.
Short answer: OpenAI began rolling out GPT-6 Astra on September 3, 2026, first to companies in its Trusted Access cybersecurity program, then to ChatGPT Plus, Pro, Business, and Enterprise plans, plus the API and AWS, within days. API pricing is $10 per million input tokens and $50 per million output tokens. Free-tier ChatGPT does not get Astra.
In my testing of the rollout notes, I cross-checked OpenAI’s own developer docs against CNBC’s reporting from launch day, because the phased-access language in OpenAI’s launch post ("rolling out today," "coming days") is vague enough that it's easy to assume you have access when you don't. I also pulled the independent benchmark numbers from ARC Prize rather than repeating OpenAI's headline score as-is, since that score comes with a footnote most coverage is skipping.
Who actually has access right now
Astra isn't a single on/off switch. OpenAI shipped it in phases, and where you land depends on your plan:
- First wave: companies enrolled in OpenAI's application-based Trusted Access cybersecurity program got Astra on day one, September 3, 2026.
- Following days: ChatGPT Plus, Pro, Business, and Enterprise subscribers, plus the API and Amazon Web Services, are getting access as the rollout expands. Pro, Business, and Enterprise plans get a higher-effort variant called GPT-6 Astra Pro.
- Enterprise workspaces: the model is off by default. A workspace admin has to turn it on before anyone on the team can use it.
- Free tier: not included. If you're on ChatGPT's free plan, my how to use ChatGPT for free guide covers what you actually get without paying — Astra isn't part of it.
The model ID on the API is gpt-6-astra, with a knowledge cutoff of April 30, 2026, a 1.05-million-token context window, and a 128,000-token output cap. It accepts text and image input and returns text only.
Step-by-step: getting on GPT-6 Astra
1. Check your plan first, not the model picker
Before you go looking for Astra in the ChatGPT interface, confirm your plan tier. Free accounts won't see it at all, and even paid accounts may not show it until the rollout reaches your cohort — OpenAI is staggering this over days, not hours.
2. Enterprise admins: turn it on manually
If your team is on Enterprise, don't wait for it to just appear. An admin has to enable Astra for the workspace in the admin console first, since it ships off by default.
3. Pick a reasoning effort level for the task
Astra exposes five effort settings — low, medium, high, xhigh, and max — instead of a single fixed reasoning depth. Low is fine for a quick rewrite; save xhigh or max for something like debugging a gnarly production issue, since higher effort costs more and takes longer.
4. Point API traffic at the new model ID
If you're calling the API directly, swap your model parameter to gpt-6-astra (or gpt-6-astra-pro where available) rather than assuming an alias will auto-upgrade your existing GPT-5.6 Sol calls. Test on a low-stakes endpoint first — the pricing and token limits are different enough to change your bill.
5. Budget for the price jump before you flip the switch
Astra's list price is noticeably higher than Sol's. Run your typical monthly token volume through the new numbers below before you migrate a production workload wholesale.
GPT-6 Astra vs. GPT-5.6 Sol vs. Claude Opus 5
Here's where the "most intelligent model ever" framing needs a second look. ARC Prize — the independent group that runs the ARC-AGI-3 benchmark — published its own numbers separately from OpenAI's launch materials, and they don't match the headline figure:
| Model | ARC-AGI-3, standard harness | ARC-AGI-3, OpenAI's provider-adapter harness | API price (input / output per 1M tokens) |
|---|---|---|---|
| GPT-6 Astra | 62.7% | 99.9% | $10 / $50 |
| Claude Opus 5 | 30.2% | N/A — harness not offered | Not the focus of this piece |
| GPT-5.6 Sol | 7.8% | N/A — harness not offered | $5 / $30 (list) |
The 99.9% score OpenAI is quoting everywhere came from a custom "provider adapter" harness that lets the model carry reasoning state between requests and compact long conversations — a setup ARC Prize ran at roughly $19K in compute, versus about $26K for the standard harness that produced 62.7%. Both numbers are real and both are published by ARC Prize; they're just measuring different things, and only one of them is comparable to how Sol and Opus 5 were scored. On FrontierMath Tier 4, OpenAI reports Astra at 98%, a figure I haven't seen independently re-run yet, so treat it as vendor-reported until a third party confirms it.
Example prompts to try once you have access
These are useful for figuring out whether Astra is worth the price jump for your actual workload, not just the benchmark:
- "Reproduce this bug using computer use: open the app, walk through the repro steps, and tell me exactly where it breaks."
- "Do this task at low reasoning effort first. If your confidence is under 80%, redo it at high effort and show me what changed."
- "Review this codebase change for security issues the way a penetration tester would, not just a linter."
- "Build me a spreadsheet from these three source files, and flag any row where the source data conflicts."
- "Summarize what changed between GPT-5.6 Sol and GPT-6 Astra on tasks like this one, citing your own uncertainty."
Common mistakes to avoid
The one I'd flag first: assuming the 99.9% ARC-AGI-3 number applies to however you're actually calling the model — it only applies to OpenAI's specific provider-adapter setup, and the standard-harness score of 62.7% is the one comparable to every other model's published result. Second, migrating production API traffic to gpt-6-astra without re-pricing it first; at $10/$50 per million tokens, a workload that was cheap on Sol's $5/$30 rate can double in cost overnight. Third, Enterprise teams assuming Astra just shows up — it's off by default and needs an admin to enable it. Fourth, treating "rolling out" as "available to me today" — OpenAI's own language is phased, and free-tier users don't get it at all, this launch or otherwise. Fifth, skipping OpenAI’s safety overview entirely: Astra is the first OpenAI model to hit the "Critical" cybersecurity capability threshold under the company's Preparedness Framework, which is the actual reason access is staggered and gated behind an application for some users.
Tools that make this easier
If you're deciding whether to move existing workflows onto Astra, start with what you're already running. My GPT-5.6 Sol pricing guide has the current discounted Sol rate for comparison, and accelerating GPT-5.6 Sol with Ultrafast is worth reading if speed, not raw capability, is your actual bottleneck. For everyday ChatGPT use rather than API calls, how to use ChatGPT and how to use ChatGPT Codex cover the interface changes Astra brings to the chat and coding products, and how to use ChatGPT agent mode is the closest existing guide to Astra's computer-use behavior. If you're weighing Astra against Anthropic's lineup before committing budget, Claude vs. ChatGPT and my wider best AI models roundup are more useful starting points than any single benchmark chart. And if the pricing jump has you reconsidering your AI budget generally, AI is getting way too expensive is a fair gut-check before you commit.
My take
The rollout itself was messier than OpenAI probably wanted — coverage from CNBC, Axios, and others went up citing OpenAI's own launch material before the company's official Astra page was reliably public, and the blog post briefly disappeared before going back up. That's a footnote, not a scandal, but it fits a pattern with this launch: the headline numbers (99.9% on ARC-AGI-3, "the AGI era" from OpenAI's president) are doing more work than the fine print supports. Astra genuinely improved on the standard-harness benchmark — 62.7% versus Sol's 7.8% and Opus 5's 30.2% is a real jump — and the cybersecurity safety numbers (0% scope violations versus Sol's 48% in OpenAI's own evaluation) are a legitimate improvement worth taking seriously given the model also crossed into "Critical" capability territory. I'd use Astra for a hard, well-defined task where the higher price is easy to justify. I wouldn't repoint a whole production pipeline at it based on a benchmark score that needed a $19,000 custom harness to hit.
Frequently Asked Questions
Is GPT-6 Astra free to use?
No. It's not available on ChatGPT's free plan at all. You need Plus, Pro, Business, or Enterprise, or an API key, and even then access is rolling out in phases rather than to everyone at once.
How do I get access to GPT-6 Astra right now?
If your organization is in OpenAI's Trusted Access cybersecurity program, you already have it as of September 3, 2026. Everyone else on a paid ChatGPT plan or the API gets access as the phased rollout reaches their account over the following days — check your plan's model picker rather than assuming it's there.
What does GPT-6 Astra cost through the API?
$10 per million input tokens and $50 per million output tokens, with cached input at $1 per million and cache writes at $12.50 per million. That's up from GPT-5.6 Sol's standard $5/$30 rate, so re-check your token volume before migrating a production workload.
Is the 99.9% ARC-AGI-3 score real?
The number itself is real and published by ARC Prize, but it came from a custom OpenAI harness that cost roughly $19,000 to run and isn't comparable to how other models were scored. The comparable figure, using the same standard harness every other model was tested with, is 62.7%.
Should I switch from GPT-5.6 Sol to GPT-6 Astra today?
For hard, well-defined tasks where accuracy matters more than cost, yes — the standard-harness benchmark gains are real. For high-volume routine workloads, I'd hold off migrating everything until you've priced out the token-cost jump against Sol's discounted rate.