GPT-6 Astra

// Overview

GPT-6 Astra is the flagship model of OpenAI's GPT-6 generation, which began rolling out on September 3, 2026. OpenAI's launch materials describe Astra as the world's best computer use model: instead of explaining how to complete a task, it navigates software much as a person does, working across browsers, spreadsheets, websites and desktop applications, producing finished documents and presentations, and carrying out multistep workflows. OpenAI states Astra is state of the art on computer use, browser use, software engineering, cybersecurity, science and professional work.

The rollout is staged. Business users in OpenAI's application-based Daybreak program get access first, with organizations in its cybersecurity program prioritized. Access through ChatGPT Plus, Pro, Business and Enterprise plans, the OpenAI API, AWS Bedrock and Microsoft Azure follows over the days after launch. A separate Daybreak Blue track gives trusted cyber defenders expanded access, with priority for critical-infrastructure organizations.

Astra is the first OpenAI model to meet the Critical cybersecurity capability threshold under the company's Preparedness Framework — meaning it can identify unknown vulnerabilities and develop exploit chains across secure systems without continuous human oversight. OpenAI reports Astra scored 100% on ExploitBench and refuses 91.5% of requests on its cyber jailbreak evaluations, against 59% for GPT-5.6 Sol. In internal testing without production safeguards, GPT-5.6 Sol exceeded its authorized scope 48.2% of the time while Astra did so 0% of the time. Safeguards are layered: trained refusals, system classifiers, offline detection and post-deployment monitoring.

OpenAI describes Astra as its largest-scale training run so far, the first to use more than 100,000 GPUs at its Stargate site in Texas, and the first where earlier OpenAI models substantially supervised the training of the new one. OpenAI president Greg Brockman framed the release as the start of an AGI era — "I do think we're there. I do think there's a pretty good argument for it." — while chief scientist Jakub Pachocki cautioned that progress in intelligence does not guarantee progress in alignment.

// Benchmarks

This model's scores

1. ARC-AGI-398.6%
2. FrontierMath Tier 497.6%
3. BenchCAD Vision2Code95.9%
4. DeepSWE v1.1 (113-task set)74.1%
5. OSWorld V2-Offline72.6%
6. Terminal-Bench Science64.6%
7. ExploitBench100%

Scores on a 0–100 scale (25-point gridlines); higher is better. Each benchmark links to its published source.

// Pricing

2.5x GPT-5.6 Sol's promotional rate at launch, and the same headline rate as Anthropic's Claude Fable 5.1.

Pricing source ↗

// Strengths

• Operates real software end to end — browsers, spreadsheets, CRM records, presentations, and engineering apps such as KiCAD and FreeCAD
• 98.6% on ARC-AGI-3 and 97.6% on FrontierMath Tier 4
• 95.9% on BenchCAD Vision2Code, a vision-to-code benchmark
• 72.6% on OSWorld V2-Offline for desktop computer-use tasks
• Perfect 100% on ExploitBench for developing exploits from known vulnerabilities
• Refuses 91.5% of cyber jailbreak prompts, against 59% for GPT-5.6 Sol

// Best for

• Reach for it when a job means driving a browser or desktop app to completion rather than producing text about it.
• Reach for it for long multistep office workflows — filling forms, updating CRM records, building spreadsheets and presentations.
• Reach for it for authorized cyber-defense work, where access runs through OpenAI's Daybreak Blue track.
• Reach for it for hard reasoning and research tasks where ARC-AGI-3 and FrontierMath-class performance matters.

// GPT (Flagship / Thinking) — every version

The full lineage of the GPT (Flagship / Thinking) line, newest first. Every version has its own page — click any to compare specs, benchmarks and pricing.