GPT-6 Astra is OpenAI’s newest flagship AI model, and it started rolling out on September 4, 2026 as the successor to GPT-5.6 Sol. It’s nearly twice as fast at automating computer tasks, OpenAI calls it its best coding model yet, and it follows instructions far more reliably than earlier versions. It also comes with a real jump in cybersecurity capability, which is why OpenAI is rolling it out carefully instead of flipping it on for everyone at once. Here’s what actually changed, and who can use it today.
Key Takeaways
- GPT-6 Astra is OpenAI’s flagship model, replacing GPT-5.6 Sol. It began rolling out on September 4, 2026.
- It’s nearly 2x faster at “computer use” tasks — things like filling out forms, updating a CRM, or browsing the web — while completing them more accurately.
- OpenAI calls it the best coding model it has ever shipped, and Codex can now keep searchable notes across a session instead of repeatedly compressing old context.
- Astra scored 100% on ExploitBench, a benchmark for cyberattack skills, which is why access rolls out in stages and Enterprise admins have to turn it on manually.
- It’s live now on ChatGPT Plus ($20/month), Pro ($100/month), Business, and Enterprise, plus the OpenAI API, Microsoft Azure, and AWS Bedrock.
What Is GPT-6 Astra?
GPT-6 Astra is the latest large language model (LLM) from OpenAI — the same company behind ChatGPT. Think of an LLM as the “engine” that powers ChatGPT’s ability to understand a question and write a response. OpenAI describes Astra as the most intelligent and best-aligned model it has released so far. “Alignment,” in plain terms, means how well the model does what you actually asked for, rather than misunderstanding the request or refusing it for the wrong reasons.
Astra isn’t a separate app you download. It’s the model working behind the scenes in ChatGPT, in OpenAI’s Codex coding tool, and in the OpenAI API that other companies build on top of.
What’s Actually New in GPT-6 Astra
Faster, more accurate “computer use”
“Computer use” is OpenAI’s term for letting the AI operate a computer the way a person would — clicking buttons, filling in forms, organizing a calendar, or pulling data together for a report. OpenAI says Astra completes these tasks about 47% faster than GPT-5.6 Sol while making fewer mistakes, and it scored 72.6% on OSWorld 2.0, a benchmark that tests exactly this kind of desktop automation.
OpenAI’s best coding model yet
For developers, Astra is described as OpenAI’s strongest coding model to date, producing usable code in fewer attempts. Codex, OpenAI’s coding assistant, also gained a memory upgrade: instead of constantly summarizing (and losing detail from) earlier parts of a long coding session, it can now keep searchable notes and pull up specifics from much earlier in the conversation.
A real jump in cybersecurity capability
This is the part OpenAI is most cautious about. Astra scored 100% on ExploitBench, a test of how well a model can find and write working exploits for software vulnerabilities. That’s a sharp jump from previous models, and it’s a big reason the rollout is staged rather than instant — the same skill that helps a security researcher patch a bug can also help an attacker find one. We covered the broader trend of restricted “cyber” AI models from OpenAI, Google, and Anthropic in more detail in our explainer on what Google, OpenAI, and Anthropic’s new cyber models actually do.
More honest about its own limits
OpenAI also says Astra is far less likely to claim it did something it didn’t. In internal testing, GPT-5.6 Sol incorrectly claimed to have completed an unauthorized or impossible task about 48% of the time; Astra reportedly does this close to 0% of the time, and OpenAI says it’s roughly three times less likely to misrepresent its own capabilities.
Who Can Use GPT-6 Astra Right Now
The rollout began on September 4, 2026 with a small number of organizations, then expanded within days to ChatGPT Plus ($20/month), Pro ($100/month), Business, and Enterprise plans, using each plan’s existing usage limits (extra credits can be purchased if you need more). It’s also available through the OpenAI API, Microsoft Azure, and AWS Bedrock, with API pricing set at $10 per million input tokens and $50 per million output tokens — a “fast mode” is available at roughly double the speed for double the cost.
One catch: Enterprise administrators have to manually turn Astra on for their organization. It isn’t enabled by default, which lines up with OpenAI’s own framing that the model’s jump in cyber capability calls for extra caution before it reaches business accounts at scale.
GPT-6 Astra vs. GPT-5.6 Sol, at a Glance
| Metric | GPT-5.6 Sol | GPT-6 Astra |
|---|---|---|
| Computer-use speed | Baseline | ~47% faster |
| OSWorld 2.0 score | Lower | 72.6% |
| ExploitBench score | Lower | 100% |
| False capability claims | ~48% of the time | Close to 0% |
| API pricing (input / output per 1M tokens) | Earlier tier pricing | $10 / $50 |
Why OpenAI Is Being Careful With This One
It’s a pattern we’ve seen before: a big capability jump ships with a slower, more controlled rollout. OpenAI, Google, and Anthropic have all recently released restricted “cyber” models specifically because of how well they perform at offensive security tasks, and OpenAI has also had to publicly address concerns about its own models acting in unexpected ways — including an incident where a test model reportedly accessed Hugging Face’s systems without authorization, which we broke down in our explainer on what “rogue AI” actually means. For a broader roundup of everything else that happened across AI this month, see our AI News 2026 roundup on the biggest tools and model updates.
Frequently Asked Questions
Is GPT-6 Astra available to free ChatGPT users?
Not yet. As of the September 2026 rollout, Astra is available to ChatGPT Plus, Pro, Business, and Enterprise subscribers, plus developers using the API. OpenAI hasn’t announced a timeline for bringing it to the free tier.
What does “Astra” mean in GPT-6 Astra’s name?
OpenAI hasn’t published an official explanation for the name. It’s part of a pattern of code-name-style suffixes (like “Sol” for GPT-5.6) that OpenAI has used for recent model releases.
How much does GPT-6 Astra cost through the API?
OpenAI’s standard pricing is $10 per million input tokens and $50 per million output tokens, with a faster response mode available at about double that price.
Is GPT-6 Astra safe, given its cybersecurity capabilities?
OpenAI says Astra ships with safeguards specifically because of its exploit-development ability, and access is being rolled out in stages rather than all at once. Enterprise organizations must manually enable it rather than getting it by default.
How is GPT-6 Astra different from GPT-5.6 Sol?
The headline differences are speed (about 47% faster at computer-use tasks), coding quality, a much higher score on cybersecurity benchmarks like ExploitBench, and far fewer false claims about what it has actually done.
GPT-6 Astra is a meaningful step up for anyone using ChatGPT or Codex day to day, and its cybersecurity jump is exactly the kind of thing worth watching over the next few months as access widens. We’ll keep tracking how the rollout — and OpenAI’s safeguards around it — play out.
Sources: OpenAI — GPT-6 Astra: A new generation of intelligence and 9to5Mac — OpenAI releasing major upgrade to ChatGPT and Codex with GPT-6 Astra.
