OpenAI has opened GPT-6 Astra to paying subscribers and the API in a rollout that began on Friday 4 September, one day after unveiling the model. The company said on X the model “is now available to all Pro, Enterprise, and Business Premium users in ChatGPT Work and Codex,” with Plus and standard Business access to follow within days.
The company bills Astra as state of the art in computer use, browsing, software engineering, cybersecurity, science and professional work. It scores 99.9 per cent on ARC-AGI-3, 97.6 per cent on FrontierMath Tier 4, 96 per cent on GPQA Diamond and 100 per cent on ExploitBench, and reaches 72.6 per cent on the OSWorld 2.0 computer use benchmark while completing tasks 47 per cent faster than GPT-5.6 Sol.
Speed is a large part of the pitch. Astra finishes browser tasks 1.9 times faster than the current Sol experience on Mind2Web, runs desktop automation in the background, and supports context windows up to a million tokens, scoring 100 per cent on 8-needle retrieval at 256,000 to 512,000 tokens and 96.3 per cent from 512,000 to a million. It also posts 57.9 per cent on Terminal-Bench 4.0 and 95.9 per cent on BenchCAD.
The working style changes as much as the scores. Astra runs multi-step workflows asynchronously, asking clarifying questions while it keeps working on the parts it can progress, and OpenAI demonstrates it on everyday office tasks from filling forms and updating customer records through to building spreadsheets, presentations and websites.
Codex picks up the most visible workflow change. Instead of compressing a session as it grows, the coding agent now keeps searchable notes that carry accumulated project detail across context windows, building on the general work agent OpenAI shipped with GPT-5.3 Codex.
API pricing sits at US$10 per million input tokens and US$50 per million output tokens, with a fast mode running at twice the speed for twice the price. The model is also live on Microsoft Azure and AWS Bedrock, and a heavier Astra Pro variant is available to Pro, Business and Enterprise users.
Early partner reports focus on the agentic side. Greg Kamradt of the ARC Prize Foundation said that “On ARC-AGI-3, Astra surpassed our human action-efficiency baseline on 96% of levels…” while Niko Grupen, head of applied research at legal AI firm Harvey, said “Astra stood out by approaching legal work the way a discerning lawyer does…”
Developer-facing testers echo the theme. John Crepezzi, who works on AI assistants at trading firm Jane Street, said “GPT-6 Astra communicates in a way that’s easier for developers to follow…” and Higgsfield AI chief executive Alex Mashrabov said “Astra gives us a significant advantage in both capability and efficiency…”
The cybersecurity results cut both ways. Astra is the first OpenAI model to meet the Critical threshold for cyber capabilities under the company’s Preparedness Framework, and OpenAI has paired the release with refusals on tasks such as exploit and malware creation, hardened jailbreak resistance and expanded monitoring. Misalignment monitoring runs in production with automatic stops when agent behaviour drifts.
Misalignment safeguards get their own numbers. OpenAI reports a 0 per cent circumvention rate on its automatic review system, against 48 per cent for GPT-5.6 Sol without safeguards, and a 2.4 per cent failure rate on internal computer-use safety benchmarks compared with Sol’s 22 per cent. The company says the model is also three times less likely to misrepresent its own capabilities, and eligible API customers can run it under Zero Data Retention.
Enterprise administrators keep the off switch. Access defaults to disabled for Enterprise organisations at launch, and OpenAI plans a program called Daybreak to steer the model’s cyber capability toward defensive security work.
The staged rollout finishes with the biggest group. Plus and standard Business subscribers get GPT-6 Astra over the days following the Friday 4 September rollout start, completing availability across every paid tier.
