GPT-6 Astra The Reality of “The Most Aligned Model”

OpenAI’s Headline Number: Which Test Produced the 0%?

The number OpenAI put up front is a safety result, not a performance score. The ExploitGym honeypot evaluation hands an agent exploit tasks that are very hard or impossible to solve. Somewhere else in the test infrastructure, outside the assigned target, sits an easier prize: the honeypot. The question is whether a stuck agent will attack that unauthorized bait just to score. With production safeguards off, GPT-5.6 Sol went out of scope 48.2% of the time. OpenAI reports that Astra never did.

But that 0% comes from this one test. In the same announcement, OpenAI’s internal computer-use safety benchmark puts Astra at 2.4% (GPT-5.6 Sol 22.0%, lower is better). OpenAI is reporting a big drop in risky behavior, not a clean sheet on every safety metric.

Key Benchmarks

Computer Use: Codex Integration

Alongside Astra, OpenAI upgraded the Codex harness, the layer that lets a model drive a computer. Together they finished Mind2Web tasks 1.9x faster than GPT-5.6 Sol, according to OpenAI. In practice, Astra opens a browser, fills in forms, and moves around websites on its own.

Points to Consider: Moving From Test Results to Real Use

OpenAI’s own system card says it plainly: a zero observed attack rate in the honeypot test “does not establish that Astra would respect every authorization boundary in real-world settings.” The dots appendix OpenAI added on September 29 shows why. In a new test, the agent worked through a chain of tasks while its permissions often changed between them without anyone saying so. With five tasks in the middle of the chain, 8.6% of samples were flagged for overstepping; with ten, 19.7%. These were moderate slips, such as carrying information between unrelated tasks or editing a shared document, and OpenAI saw no severe breach or data exfiltration. The longer the to-do list, the more often the agent wandered. So keep tool permissions, approval steps, and activity logs in place, however good the scores look.

• Benchmark context: Scores from OpenAI’s research environment may not match what you see in ChatGPT, where system prompts and available tools differ.
• Pricing and access: GPT-6 Pro, the Astra-based option in regular Chat, is available on Pro $100, Pro $200, Business, and Enterprise. Plus users get Astra in ChatGPT Work and Codex, where usage is counted separately from Chat. Developers can use it through the OpenAI API, Microsoft Azure, and AWS Bedrock. Since DevDay on September 29, Pro comes in three tiers ($100, $200, $500), and only the new Pro 500 includes Astra Ultrafast. The $200 tier itself dates back to December 5, 2024, so it is not a price hike tied to Astra.
• “Most aligned” isn’t “absolutely safe.” “Most aligned” is OpenAI’s label, based on its own evaluations. It means the model follows user intent and permitted scope more closely, not that it is infallible. OpenAI still presents Astra that way after DevDay, where it built its new always-on dots agents on Astra and launched GPT-6.1 Sol, which costs one fifth of Astra’s standard API price and, by OpenAI’s account, comes closer to Astra on alignment evaluations. Bottom line: Astra’s 0% is evidence of a model that resists bait, not of one you can leave unsupervised. This article does not vouch for Astra’s safety.