GPT-6 Astra: what actually changes in ChatGPT and ChatGPT Work

The “GPT-6” button you’re hunting for doesn’t exist

OpenAI announced GPT-6 Astra on 3 September 2026, and the first new feature is a vocabulary lesson. Astra is a model, ChatGPT is a product, and LoKan Sardari pointed out on 5 September that the second doesn’t put the first in front of you the same way everywhere: in Work and in Codex, the Astra name is visible and selectable; in plain old Chat, the top-shelf version it powers is called GPT-6 Pro and stays locked behind the pricier plans. It surfaces when you crank the “Thinking Effort” dial to maximum, which is a fairly elegant way of putting thought on the meter.

On availability, the stories already disagree. OpenAI’s announcement describes a rollout that starts with a limited number of organisations before reaching, over the following days, every Plus, Pro, Business and Enterprise subscriber, plus the API, Azure and AWS Bedrock. ZDNET wrote on 4 September that it was available “from today”. The spec sheet: 1,050,000 tokens of context, 128,000 tokens of output, text and images in, text out, and internal knowledge that stops on 30 April 2026.

What it does for you, in practice

OpenAI’s list of tasks is reassuringly dull: filling in online forms, updating customer records in a CRM, sorting out a calendar, running a piece of research and writing the summary straight into your inbox or your document editor. Then the ambition creeps up: analysing data, generating charts, building a website, running front-end QA checks to confirm the features actually work, installing and testing software.

The headline numbers are mostly about speed. On OSWorld 2.0, Astra hits 72.6% in roughly 40 minutes per task, against 65.7% in 75 minutes for GPT-5.6 Sol. The Codex harness got an update alongside it, which yields completion 1.9 times faster than the current Sol experience on Mind2Web. On Agents’ Last Exam, which tests professional tasks inside real software, Astra scores 59.3% against 55.5% for Claude Opus 5 and 53.6% for Sol, while burning roughly 65% fewer output tokens.

One detail the press release doesn’t lead with: the model on its own touches nothing. It doesn’t edit a local file, doesn’t fill in a form, doesn’t build a deck. Work and Codex are what hand it the tools. Astra is the brain; the hands are rented.

The first model rated “Critical” for cybersecurity

This is the part where nobody’s joking. For the first time in the history of OpenAI deployments, a model has reached the “Critical” capability threshold for cybersecurity under the Preparedness Framework. In the lab, stripped of its production guardrails, Astra found zero-day vulnerabilities unaided and assembled working exploit chains against browsers and operating systems. It scores 100% on ExploitBench, against 78.5% for Sol, and solves 88% of SRE-Bench’s reverse-engineering tasks on the first attempt. It’s also rated “High” risk in biology and chemistry.

The trade-off: Astra refuses to produce proof-of-concept exploits for the general public, those capabilities being reserved for vetted participants in the Trusted Access for Cyber programme, codenamed Daybreak Blue.

More obedient, harder to read

On behaviour, OpenAI built an evaluation inspired by the Hugging Face incident, measuring whether a model handed a hard or impossible task wanders outside its remit. Sol did so 48% of the time without production guardrails; Astra, 0%. False statements about its own work are down fourfold, and across 54,000 simulated internal engineering tasks, reports of high-severity misaligned behaviour fall by 53%.

But the System Card flags a drop in chain-of-thought monitorability: the model now reasons further ahead without saying any of it out loud. The API, for its part, bills $10 per million input tokens and $50 per million output.

So you get a colleague who lies four times less, never once oversteps the brief, and increasingly keeps its thinking to itself. Three for three, as long as you stop counting at two.