OpenAI's GPT-6 Astra Is First Model Rated 'Critical' Cyber Risk

OpenAI began rolling out GPT-6 Astra on September 3, 2026, first to a limited set of organizations before wider access lands across ChatGPT Plus, Pro, Business, and Enterprise tiers, plus the API and AWS Bedrock in the following days. The model is the successor to GPT-5.6 and arrives more than a year after GPT-5, positioned as a frontier system built around real-world computer use.

That framing shows up most clearly in benchmarks. On OSWorld 2.0, a test of an AI agent's ability to operate an actual desktop, GPT-6 Astra scores 72.6%, ahead of Claude Opus 5 at 70.2% and GPT-5.6 Sol at 65.7%. OpenAI says the model can fill out forms, update CRM records, run frontend QA checks, and troubleshoot software without step-by-step human direction.

A new risk classification

Under OpenAI's Preparedness Framework, GPT-6 Astra is the first model the company has classified as Critical for cybersecurity risk. On ExploitBench it hits a perfect 100%, up from GPT-5.6 Sol's 78.5%, and it scores 42.4% on ExploitGym versus 30.3% for its predecessor. It also reaches 99.2% on SRE-Bench binary reverse engineering within four attempts, compared with 68.7% for GPT-5.6 Sol.

OpenAI reports the model refuses cyber jailbreak attempts 91.5% of the time, against 59% for the prior model, a gap the company points to as evidence that added capability hasn't come without added safeguards.

Elsewhere the numbers are more mixed. GPT-6 Astra ties GPT-5.6 Sol on the Artificial Analysis Intelligence Index at 61, trailing Claude Fable 5.1 by 5 points, even as it edges out Sol by 2 points on the Coding Agent Index at similar cost. It posts 97.6% on FrontierMath Tier 4 and 74.1% on the DeepSWE v1.1 agentic coding benchmark, a smaller margin over Sol's 70.8%.

Long-context performance stands out: with a 1.05-million-token window, the model hits 100% on OpenAI's eight-needle retrieval test between 256K and 512K tokens, dipping only slightly to 96.3% between 512K and 1M. On ARC-AGI 3, released in March 2026, it claims 99.9% accuracy and "human parity," though that figure depends on OpenAI's custom Provider Adapter harness. The default ARC-AGI harness puts it at 62.7%, a discrepancy worth noting for anyone comparing results across labs.

Pricing and access

API pricing is listed at $10 per million input tokens and $50 per million output tokens, though some coverage has shorthanded this as pricing simply "starting around $10 per million tokens," a simplification that omits the steeper output rate.

OpenAI CEO Greg Brockman offered the boldest framing of the launch, suggesting the model could mark the point at which the industry crosses into AGI territory.

The new model may constitute the turning point at which humanity has entered the era of Artificial General Intelligence.

That claim rests on OpenAI's own internal assessments and benchmark harness choices, and independent verification of AGI-level performance has not yet materialized outside the company's reporting.