OpenAI Astra Launches: Browser Control and Hidden Reasoning

OpenAI released Astra on 3 September 2026, and its headline trick is not writing - it is using a computer and a browser the way you would.

OpenAI calls Astra its most capable model so far, and its most aligned.

It is also the first one the company has had to hold back and re-test before letting anyone near it.

Here is what actually changed, in plain terms.

What Astra Actually Does

OpenAI describes Astra as "a new frontier on computer and browser use", which means it is built to click, type, scroll and fill things in rather than just answer you. Company president Greg Brockman called it "our most intelligent and, also very importantly, our most aligned model yet".

It is also being sold as the strongest coding model to date, scoring above OpenAI's own Sol and Anthropic's Fable on bug-finding and terminal work.

Access started with OpenAI's Daybreak security customers, with Pro, Plus, Enterprise, Business and the API following over the following week. The most powerful security features are being handed out more slowly, to a small group of testers first.

The Security Part Is the Real Story

Astra is the first model OpenAI has ever rated Critical for cybersecurity under its own Preparedness Framework. In its own words, that means it can find security flaws nobody knew about and build working attacks on well-defended systems, without a person guiding each step.

During internal testing it scored 100 percent on one exploit benchmark and turned up two genuine unknown vulnerabilities along the way, which OpenAI says it is reporting to the maintainers.

OpenAI says it trained Astra to refuse this kind of request far more reliably than before - 91.5 percent of its cyber jailbreak tests, against 59 percent for the previous model.

Why Researchers Are Uneasy

Astra uses a technique reported as "opaque recurrence", or recurrent depth, where the model loops over a question internally instead of thinking in a readable line of steps.

That matters because the readable line, the chain of thought, is how researchers check why a model did something. Less of it in writing means less to inspect.

Redwood Research chief executive Buck Shlegeris said he was "extremely concerned" by the reports, and his colleague Ryan Greenblatt warned about models eventually reasoning almost entirely out of sight. OpenAI chief scientist Jakub Pachocki has pushed back, saying legible reasoning remains a core goal, while conceding that "as model capabilities are increasing, monitorability is getting more challenging".

OpenAI says Astra's use of the technique is limited and its chain of thought stays readable.

What Changes for You

Not much this week, and that is the honest answer.

If you use the ChatGPT app or a paid plan, the model underneath will change and you will mostly notice better answers on longer, fiddlier tasks.

The one thing worth expecting is interruptions. OpenAI has said outright that its new safety checks can slow, pause or stop legitimate work, that ChatGPT and Codex may ask you to approve an action before continuing, and that on the API the task simply stops.

If browser automation is the part you want, note that ChatGPT Atlas was retired in August 2026 and its browsing features moved into the desktop app, so there is no separate browser to chase. Comet Browser is the maintained alternative if you want an assistant living inside the page.

Quick questions

The model changed. The app on your phone did not.

Download ChatGPT for Android and see whether the answers feel any different this week, or tell us what you make of Astra if you have already been given access.