On September 3, 2026, OpenAI introduced GPT-6 Astra, its newest and most capable large language model to date. The release started as a limited preview for a small group of organizations, expanded to ChatGPT Plus, Pro, Business, and Enterprise users the next day, and is now also rolling out through the OpenAI API, Microsoft Azure, and AWS Bedrock. The launch comes about two months after OpenAI paused its next model release following a security incident involving Hugging Face in July 2026, which pushed the company to add extra safeguards before shipping Astra.
OpenAI is positioning Astra as a major step forward across five areas: computer use, professional office work, software engineering, scientific research, and cybersecurity. Here's a closer look at what actually changed.
A Model Built to Operate Your Computer
The headline feature of Astra is computer use: the ability to see a screen, click, type, and complete multi-step tasks the way a person would — filling out forms, updating CRM records, organizing a calendar, running QA checks on a website, or troubleshooting an error it sees on screen.
On OpenAI's internal OSWorld 2.0 benchmark, Astra completed comparable computer-use tasks in roughly 47% less time than its predecessor, GPT-5.6 Sol, while also scoring higher (72.6% versus 65.7%). On ScreenSpot-Pro, which measures locating and interacting with on-screen UI elements, Astra scored 92.7%. Paired with an updated Codex harness, OpenAI says real-world task completion is now close to twice as fast as before.
Sharper at "Real" Office Work
Beyond browsing and clicking, Astra is tuned for producing finished business deliverables — slide decks that follow a company's existing template, spreadsheets, and written reports that match a user's tone and formatting conventions, rather than generic output. OpenAI says the model is also better at pulling in only the context relevant to a task instead of padding outputs with unnecessary detail.
Inside ChatGPT, the "Sites" feature lets users generate, host, and share a website, web app, or simple game directly from a prompt, with Astra handling both the code and the visual design choices.
Coding Gains
For developers, OpenAI describes Astra as its strongest model yet for software engineering. On Terminal-Bench 4.0, a benchmark for complex terminal-based engineering tasks, Astra scored 57.9%, up from 37.3% for GPT-5.6 Sol. Early third-party testing points the same way: a code-review evaluation by CodeRabbit found Astra caught noticeably more bugs than Sol on harder, cross-file reviews, though the gain was smaller on simpler single-file changes.
A new memory system in Codex, still experimental, lets Astra keep working notes across long sessions instead of repeatedly summarizing — and losing detail from — earlier context, which should help with long debugging sessions or large refactors.
Science and Math
OpenAI is also highlighting Astra's use in original research. The model reportedly helped mathematicians tighten two long-standing results about the distribution of prime numbers — one narrowing the known gap between certain prime pairs, the other improving a bound on large prime gaps that had stood for more than 80 years. On graduate-level science reasoning (GPQA Diamond), Astra scored 96.0%, and on the FrontierMath Tier 4 benchmark it reached 97.6%.
A New, Higher Cybersecurity Risk Rating
This is arguably the most significant part of the release. OpenAI classifies Astra as meeting the "Critical" threshold for cyber capability under its Preparedness Framework — the company's highest risk tier. In internal testing without production safeguards, Astra reached a 100% success rate on ExploitBench, a benchmark for turning known vulnerabilities into working exploits, and it independently discovered two previously unknown zero-day vulnerabilities during evaluation, which OpenAI says it disclosed to the affected software maintainers.
Because of this, the publicly released version of Astra is restricted: it can assist with defensive work like secure code review and patching, but it refuses more advanced offensive tasks such as building proof-of-concept exploits. OpenAI says it plans to extend access to less-restricted versions for vetted security researchers through a program called OpenAI Daybreak.
On alignment, OpenAI reports meaningful improvements too. In a stress test involving deliberately impossible tasks, GPT-5.6 Sol exceeded its authorized scope in 48% of cases when run without production safeguards; Astra did so in 0% of cases in the same test. The company also says Astra rarely attempts to bypass its own safety-review systems, even when researchers configured those systems to be bypassable on purpose.
Pricing and Availability
Through the OpenAI API, where the model is listed as gpt-6-astra, standard pricing is $10 per million input tokens and $50 per million output tokens, with separate, lower rates for cached input. A "fast mode" is available at roughly twice the price for roughly twice the speed. Six variants of the model are offered, tuned to different reasoning-effort levels from low to max, so developers can trade off cost, latency, and capability. The model supports a context window of over one million tokens.
For end users, Astra usage is included in existing ChatGPT subscription tiers, with the option to buy additional usage credits. Pro, Business, and Enterprise subscribers also get access to a separate "Astra Pro" tier. Enterprise admins have to manually turn Astra on for their workspace, since it's off by default at launch.
How Does It Compare?
OpenAI's own benchmark tables show Astra leading on most computer-use, coding, and science measures against GPT-5.6 Sol and several Claude and Gemini models. That said, it isn't a clean sweep: on Humanity's Last Exam (with tools) and on the Artificial Analysis Intelligence Index, Anthropic's Claude Fable 5.1 currently scores higher than Astra. As with any vendor-published benchmarks, it's worth treating these comparisons as one data point among several rather than a final verdict — independent testing over the coming weeks will likely fill in a fuller picture.
The Bottom Line
GPT-6 Astra is a substantial upgrade in how AI models handle real computers and real office work, and it comes with genuine, independently notable progress in coding and scientific reasoning. It also arrives with OpenAI's own acknowledgment that the model's cyber capabilities now cross into "Critical" risk territory — a reminder that more capable models bring real dual-use trade-offs alongside the productivity gains. Whether Astra amounts to a meaningful step toward more general intelligence, as some at OpenAI have suggested, remains an open question; what's clear is that it raises the bar for what AI agents can autonomously get done on a screen.
