GPT-6 Astra: Release Date, Pricing, Benchmarks, and Rollout (2026)

OpenAI shipped GPT-6 Astra on September 3, 2026, four weeks after saying it had slowed the model over cyber risk. Here’s what’s confirmed, what’s still rolling out, and what it costs.

The most-searched unreleased model of the year is now a released one. OpenAI announced GPT-6 Astra on September 3, calling it “the most intelligent and aligned model in the world,” with an API model ID of gpt-6-astra, a 1M-token context window, and a price 2.5x GPT-5.6 Sol's. It is rolling out in stages, with the cyber-sensitive capabilities gated behind a trusted-access program.

This page was a rumor tracker until September 3. It's now the launch breakdown, and it will keep updating as the rollout completes.

TL;DR

• GPT-6 Astra launched September 3, 2026. The name is confirmed; the API model ID is gpt-6-astra
• Rollout is staged: a limited set of organizations on day one, then ChatGPT Plus, Pro, Business, and Enterprise “over the coming days,” plus the OpenAI API and AWS. Enterprise access is off by default until an admin enables it
• API pricing is $10 per million input tokens and $50 per million output, with cached input at $1, batch at half price, and a Fast mode at 2x the rate. That's 2.5x GPT-5.6 Sol's current promotional rate
• Specs: 1,050,000-token context, 128K max output, text and image input, knowledge cutoff April 30, 2026
• OpenAI says Astra meets the “Critical” cybersecurity threshold under its Preparedness Framework. The public model refuses advanced cyber tasks like proof-of-concept exploits; looser safeguards go to vetted organizations through the Daybreak program
• Benchmarks are OpenAI’s own: Terminal-Bench 4.0 57.9% (Sol 37.3%), DeepSWE v1.1 74.1% (Sol 72.7%), OSWorld 2.0 72.6% (Sol 65.7%), FrontierMath Tier 4 97.6% (Sol 83.0%)
• Building today: Qwen, GLM, DeepSeek, and Kimi are live on Yotta AI Gateway behind OpenAI-compatible endpoints, with Claude being added, so slotting Astra into a routing layer next to them is a config change

What OpenAI Actually Shipped

Three things stand out in the announcement. First, the positioning is computer use and software engineering, not chat: OpenAI calls Astra “a new frontier in the speed, accuracy, and safety of computer use” and “the best model for software engineering to date,” and claims 1.9x faster task completion than Sol on the Mind2Web benchmark. Second, the cyber story that delayed the launch is now the headline feature, with a 100% score on ExploitBench against Sol’s 78.5% and a model that OpenAI says can find and chain zero-days on its own. Third, the alignment numbers moved as much as the capability numbers: OpenAI reports a hallucination benchmark of 4.2% versus Sol’s 12.2%.

Codex gets a companion update: a note-keeping feature that preserves context across context windows instead of compressing earlier work into summaries. It ships as experimental and becomes the Astra default “in the coming weeks.”

Rollout: Who Gets It When

If you don’t see it in your account or API tier yet, that’s the schedule, not a problem. This page will update when general availability is confirmed.

Pricing

Two and a half times the previous flagship on both input and output, against Sol's current promotional rates. For cost context against the open field: Qwen 3.8-Max, DeepSeek V4 Pro, GLM 5.3, and Kimi K3 all price well under Astra’s output rate, in most cases by an order of magnitude. Different capability class, different jobs, but the spread is exactly why multi-model routing exists. For the full rate card, cached and batch tiers, and worked cost math, see GPT-6 Astra pricing.

Specs

Audio and video input are not supported at launch.

Benchmarks, With the Usual Caveat

Every number below is OpenAI’s, published at launch, with no independent replication yet. Validate on your own workload before it changes a procurement decision.

The pattern: large jumps on terminal, computer-use, math, and cyber tasks; small ones on the saturated benchmarks like GPQA and DeepSWE. On DeepSWE specifically, the 74.1% is a point and a half over Sol and within striking distance of what Z.ai reports for GLM 5.3 (66.9%) at a fraction of the price, though those are two vendors’ runs of the same test, not a head-to-head.

The Cyber Gate

This is the part that delayed the launch, and it’s still shaping the rollout. OpenAI says Astra is the first model to meet its “Critical” cybersecurity threshold, meaning it can identify previously unknown vulnerabilities and build working exploits without step-by-step human guidance. During testing it found and chained two zero-days, which OpenAI disclosed to the maintainers.

The public model is trained to refuse advanced cyber tasks such as writing proof-of-concept exploits. Less restrictive safeguards go to vetted organizations through OpenAI Daybreak, which OpenAI says will expand “in the coming weeks.” If your use case is security research, that program is the door; the general model will say no.

The Washington Wrinkle

Worth keeping in mind for the next one: OpenAI’s flagship releases now pass through a government review step. The US administration publicly weighed in on GPT-5.6’s release timing earlier this summer, Altman briefed officials before that rollout, and the Astra delay was framed around a safety threshold rather than a schedule. Whatever you think of that arrangement, it means OpenAI launch dates are now a political calendar as much as a technical one, and specific predictions for the next model deserve the same skepticism the GPT-6 rumors did.

What to Build On

The practical takeaway hasn’t changed with the launch. It’s still “don’t architect around one vendor’s roadmap,” it’s just that the roadmap now has a $50-per-million entry on it.

Astra is a strong candidate for the hardest tier of agentic and engineering work, at a price that makes routing the rest of your traffic elsewhere the obvious move. The current open field, Qwen 3.8-Max, DeepSeek V4 Pro and V4 Flash, GLM 5.3, Kimi K3, covers the everyday middle at a fraction of the cost, and the architecture decision that survives every launch is an OpenAI-compatible endpoint you can repoint. We covered the mechanics in how to switch models without changing your code and the field in the best OpenAI API alternatives.

That’s the design Yotta AI Gateway is built for: Qwen, GLM, DeepSeek, Kimi, and the rest of the catalog behind one API key with OpenAI-compatible endpoints, with Claude being added. Astra becomes one more routing option for the tail, not a migration.

Frequently Asked Questions

When was GPT-6 released? September 3, 2026, as GPT-6 Astra. It launched to a limited set of organizations first, with ChatGPT paid tiers, the API, and AWS following over the coming days.

Is Astra GPT-6? Yes. OpenAI’s official name is GPT-6 Astra, and the API model ID is gpt-6-astra. The two names refer to the same model.

How much does GPT-6 Astra cost? $10 per million input tokens and $50 per million output on the API, with cached input at $1, batch and flex processing at half price, and a Fast mode at double the rate. That's 2.5x GPT-5.6 Sol's current promotional pricing.

What is the GPT-6 Astra context window? 1,050,000 tokens, with a maximum output of 128,000 tokens and a knowledge cutoff of April 30, 2026.

Can I use GPT-6 Astra in the API today? It’s rolling out. Trusted-access organizations had it on day one; the model page says broader API access is “coming soon.” Check your account rather than assuming.

Why did OpenAI delay Astra? Internal evaluations put it past OpenAI’s “Critical” cybersecurity threshold, meaning it can find and exploit vulnerabilities autonomously. OpenAI slowed the release on August 7, added safeguards, and shipped with the advanced cyber capabilities restricted to vetted organizations.

Is GPT-6 Astra better than GPT-5.6 Sol? On OpenAI’s own benchmarks, yes across the board, with the biggest gains on terminal, computer-use, math, and cyber tasks. Those numbers have no independent replication yet.

What should I build on if I don’t want to pay $50 per million tokens? Whatever wins on your workload, behind an OpenAI-compatible endpoint so you can route to Astra for the hard tail later. Qwen, GLM, DeepSeek, and Kimi models are live on Yotta AI Gateway with exactly that interface, and Claude is being added.

Is there a GPT-6 Ultra? Not as of September 6, 2026. OpenAI has announced GPT-6 Astra and, inside ChatGPT, a GPT-6 Pro tier powered by it. No "Ultra" model or plan has been announced, and anything using that name is speculation.

Is ChatGPT 6 the same as GPT-6 Astra? Yes. "ChatGPT 6" is how people search for it, but the product is GPT-6 Astra, which appears in ChatGPT as GPT-6 Pro on paid plans. There's no separate ChatGPT 6 release.

Bottom Line

GPT-6 Astra is real, it’s out, and it’s expensive. The capability jump is concentrated where OpenAI aimed it, computer use and software engineering, and the cyber capability that delayed it is now its most-discussed feature and its most-restricted one. The rollout will take days to reach everyone and weeks to loosen.

The winning move is the same as it was when this page tracked rumors: build on the models that exist, keep the endpoint OpenAI-compatible, and treat Astra as a routing destination for the traffic that earns its price. One API key on Yotta AI Gateway is the low-drama way to stay ready for that.