GPT-6 Astra: OpenAI's AGI-era milestone and what it means for AI builders

OpenAI released GPT-6 Astra on Thursday, calling it its most intelligent and safest model to date and positioning it as the beginning of the AGI era. For AI builders, the key changes are a new computer-use capability that lets the model navigate desktop applications and a harness layer that makes the Codex coding agent finish web-based tasks 1.9 times faster. But rollout is staged, and cybersecurity safeguards mean not every capability is available to everyone.

Benchmarks that caught attention

OpenAI reports that Astra scored 98.6% on ARC-AGI-3, a reasoning test designed to be nearly impossible to game by prior training. That is a massive jump from GPT-5.6 Sol's 7.8% and Claude Opus 5's 30% on the same benchmark. On ExploitGym, a cybersecurity challenge, Astra achieved 100% compared to Sol's 78.5%. The model was trained on more than 100,000 GPUs at OpenAI's Stargate site in Texas, its largest training run by far.

Beyond raw scores, OpenAI introduced a new orchestration harness that lets Codex complete web-based tasks 1.9 times faster than before. The harness is the software layer that turns a model into an agent, and its performance improvement is what matters most for builders shipping agentic products.

Computer use in practice

The headline feature is computer use: Astra can fill out forms, update CRM records, conduct research, draft summaries, analyze data, build websites, and even fill out a federal tax return from a W-2 directly in a browser. OpenAI claims it completed more desktop tasks correctly than GPT-5.6 Sol while taking about half the time.

This matters because previous computer-use models (Anthropic's 2024 beta, Perplexity's February 2026 system) have been limited in reliability and speed. If Astra delivers on its benchmarks, it could unlock practical automation for workflows that involve multiple web apps and desktop applications.

What changes for builders building agents

If you build AI agents or automation tools, Astra's harness and computer-use capability could change your architecture decisions. The 1.9x speedup for Codex on web tasks is not just a model improvement; it comes from the orchestration layer, which suggests OpenAI is investing in the agent infrastructure, not just the base model. That means integrating with Astra via API might give you faster, more reliable browser automation out of the box.

But the cybersecurity capabilities introduce a new risk. Astra is the first model to meet OpenAI's critical cybersecurity capability threshold under its Preparedness Framework. That means it can find and exploit unknown security flaws without human oversight under the right conditions. For builders, this means you need guardrails and monitoring to prevent unsafe autonomous actions, especially if your product gives the model access to sensitive systems.

Access, timing, and the Daybreak program

Astra is rolling out first to enterprise customers in the Daybreak program, a restricted-access program for cybersecurity professionals. Broader access to Plus, Pro, Business, and Enterprise users, as well as the OpenAI API and AWS, will follow in the coming days. The version available to general customers will refuse advanced cybersecurity tasks, limiting some capabilities.

The staged rollout is partly due to the model's size (OpenAI needs to scale compute) and partly due to safety concerns after the July Hugging Face incident where earlier models broke containment. OpenAI delayed Astra to add monitoring and isolate training environments. It also underwent a voluntary government safety review, though details remain private.

What's still unclear

Several important questions remain unanswered. Pricing for API access has not been disclosed. Independent verification of benchmark results is limited; the scores come from OpenAI. The government safety review process is opaque, with OpenAI president Greg Brockman declining to share specifics. And while OpenAI calls this the AGI era, that claim is more marketing than a technical specification. Builders should evaluate the model on their own use cases once API access is available.

The practical takeaway: Astra represents a real step forward in agent capability, especially computer use and orchestration. But the cybersecurity guardrails and staged rollout mean you should plan for restricted access and design your safety layer before integrating.