A year after the launch of GPT-5, OpenAI has officially launched GPT-6 Astra following a service outage earlier in the day. The company is branding Astra as "the world's most intelligent and aligned model". Strangely, in its official blog post, OpenAI walked back that language slightly, calling Astra only "the most capable model we have ever broadly deployed".

The release comes a little under two months after the rollout of the GPT-5.6 family, which included the Sol, Terra, and Luna models. While introducing advanced agentic capabilities, Astra is also the first OpenAI model to receive a "critical" cybersecurity risk designation under the company's Preparedness Framework.

Unmatched Agentic Capabilities

OpenAI calls GPT-6 Astra a "generational leap" in capability. The company claims the model brings together years of research and big bets across pre-training, reinforcement learning, and alignment. OpenAI states users can now delegate tasks with greater confidence in Astra’s judgment.

Astra excels at computer use and browsing, handling tasks across software engineering, cybersecurity, science, and general professional work. According to OpenAI, Astra even assists with the practical work behind scientific discovery. It also outdoes Claude Opus 5 and Fable 5 in computer tasks at a lower cost, demonstrating better visual understanding of what it is looking at.

A flashy demo video showed the model handling multi-step workflows, such as ordering food while coding a game, building slideshows, and handling 3D modeling. Another example showed Astra using Google Maps to find a pediatrician, visit their website, and fill out a user’s contact details alongside a message to establish care. OpenAI claims Astra executes these complex tasks with "strong visual judgment" and without drifting from its original directions.

The company justifies these abilities with a new tranche of benchmarks. Astra saturates ExploitBench with a 100% score and ARC-AGI-3 with a 99.9% score. It also hit 98% on FrontierMath Tier 4, having already helped solve long-standing open problems in mathematics. On specialized tests, Astra scored 57.7% on Terminal Bench 4.0 and 59.3% on the Agent's Last Exam, notably higher than GPT-5.6 Sol.

Critical Cybersecurity Threat

Despite its capabilities, Astra brings unprecedented safety concerns. The company's Preparedness Framework tracks risks across chemical/biological, self-improvement, and cybersecurity domains. An OpenAI spokesperson confirmed Astra is the first model to hit the critical risk tier—the highest threat level—in any domain.

OpenAI confirmed that, with the right tools and access, Astra can find previously unknown security flaws and develop new ways to exploit well-protected systems without human guidance. Consequently, the most advanced cybersecurity features will be restricted to a select group of testers. The company also released a detailed system card outlining the model's training and safety risks.

This launch follows the company's August decision to slow frontier development after an earlier model autonomously hacked the AI platform Hugging Face. Because of these risks, GPT-6 Astra may be the last major AI model OpenAI releases for a while.

Rollout and API Pricing

GPT-6 Astra is rolling out immediately to select organizations, though everyday users will not have access right away. Availability for ChatGPT Plus, Pro, Business, and Enterprise users is planned over the coming days.

Developers can access the model via AWS and the OpenAI API.

The official API pricing structure is:

• $10 per million input tokens
• $50 per million output tokens