GPT-6 Astra went live this week as OpenAI’s newest flagship model, and the reaction has been genuinely split. On one side, OpenAI president Greg Brockman called it the start of the “AGI era.” On the other, AI safety researchers immediately flagged a real concern: Astra’s reasoning is harder to audit than previous models. Both things are true at once, and this guide covers what Astra actually does, what the benchmarks show, and how to access it.
GPT-6 Astra is OpenAI’s new flagship model, rolling out first to enterprise customers, then to all ChatGPT Plus, Pro, Business, and Enterprise users, plus the API, Azure, and AWS Bedrock. It leads on computer-use and coding benchmarks, but it’s also OpenAI’s first model to hit the “Critical” cybersecurity risk threshold under the company’s own safety framework.
What GPT-6 Astra Actually Does Differently
Astra is built around computer use: instead of just answering questions, it navigates real software the way a person would, clicking through browsers, filling out forms, and working inside spreadsheets and design tools. On OpenAI’s own OSWorld 2.0 benchmark, Astra completed tasks about 47% faster than the previous model, GPT-5.6 Sol, while also scoring higher. It also introduces a feature that lets it keep persistent notes across long coding sessions, instead of repeatedly summarizing and losing detail the way older models did.
How Astra Compares on Real Benchmarks
These figures come directly from OpenAI’s own published launch materials, comparing Astra against its predecessor and against Anthropic’s Claude models:
| Benchmark | GPT-6 Astra | GPT-5.6 Sol | Claude Fable 5.1 |
|---|---|---|---|
| Computer use (OSWorld 2.0) | 72.6% | 65.7% | — |
| Coding (Terminal-Bench 4.0) | 57.9% | 37.3% | 55.8% |
| Math (FrontierMath Tier 4) | 97.6% | 83.0% | 87.8% |
| Cybersecurity (ExploitBench) | 100.0% | 78.5% | — |
The Part of the Launch That Split Opinion
Astra uses a reasoning technique OpenAI calls “recurrent depth,” which makes the model more efficient but also makes its internal chain of thought harder for researchers to fully monitor. OpenAI’s own launch materials acknowledge this directly, stating that Astra’s written reasoning is harder to audit than GPT-5.6 Sol’s on tasks specifically designed to test evasion. At the same time, Astra also reached OpenAI’s “Critical” cybersecurity capability threshold, meaning it can, in the right hands, discover and develop working exploits for previously unknown vulnerabilities without a person guiding each step.
To be fair, OpenAI paired this with real restrictions: the publicly available version refuses to help build proof-of-concept exploits, and the company reports that Astra never attempted to bypass its own safety review system in internal testing, even when that system was deliberately configured to be bypassable. Whether that’s reassuring or exactly the kind of claim that needs independent verification is, unsurprisingly, where the online reaction has split.
Because of Astra’s cybersecurity capability level, the most sensitive defensive features are currently gated behind OpenAI’s Daybreak program rather than available to everyone at launch.
How to Access GPT-6 Astra
Step 1: Check Your ChatGPT Plan
Astra is rolling out first to a limited set of enterprise organizations, then to all ChatGPT Plus, Pro, Business, and Enterprise users over the following days. If you don’t see it yet, check back in a day or two rather than assuming something’s wrong.
Step 2: Use It Through the API
Developers can call the model directly as gpt-6-astra through the OpenAI API. Standard pricing is $10 per million input tokens and $50 per million output tokens, with a Fast mode available at double the price for roughly double the speed.
Step 3: Access It Through Azure or AWS
If your infrastructure already runs on Microsoft Azure or Amazon Bedrock, Astra is available there too, so you don’t need a separate OpenAI billing relationship.
Step 4: Enable It for Your Enterprise Workspace
Enterprise administrators need to manually turn Astra on for their workspace — it’s off by default at launch, even for eligible plans.
Pro, Business, and Enterprise users also get access to a separate “Astra Pro” tier. If a task needs deeper reasoning, check whether you’re on standard Astra or Astra Pro before assuming the model has hit its limits.
Frequently Asked Questions
Is GPT-6 Astra actually AGI?
OpenAI’s own president called it a step toward AGI, but independent reviewers have pushed back, noting Astra hasn’t been shown to outperform humans at most economically valuable work — OpenAI’s own historical bar for the term. Some of its most impressive results also depend on surrounding tools and agent infrastructure, not the raw model alone.
Why is Astra’s reasoning harder to monitor?
Its efficiency gains come partly from solving problems in fewer written steps, which leaves less of its reasoning process visible to researchers. OpenAI says it takes this seriously and continues to treat monitorability as an active research priority, not a solved problem.
Can I use Astra for cybersecurity work right now?
The publicly available version supports defensive tasks like secure code review and patching, but it refuses more advanced requests, such as building proof-of-concept exploits. Full access to those capabilities is currently limited to vetted participants in OpenAI’s Daybreak program.
ChatGPT Memory: Train AI to Remember Everything About You — another recent ChatGPT upgrade worth knowing about alongside this one.