OpenAI Launches GPT-6 Astra, Claims State-of-the-Art Computer Use and 98% on FrontierMath Tier 4

OpenAI has launched GPT-6 Astra, its newest flagship model, which the company calls "the world's most intelligent and aligned model." The release follows an outage on OpenAI's systems earlier the same day.

Astra is not yet broadly available. OpenAI says it is rolling out today to "a limited set of organizations," with general availability for ChatGPT Plus, Pro, Business, and Enterprise users promised in "the coming days." No API pricing or context window figures have been disclosed.

What OpenAI claims

According to OpenAI, GPT-6 Astra represents a "generational leap" built on advances across pre-training, reinforcement learning, and alignment. The company's stated benchmark results include:

• FrontierMath Tier 4: 98%, which OpenAI says is a saturation of the benchmark. The company claims Astra has already helped solve unspecified "long-standing open problems" in mathematics.
• ARC-AGI-3: 99.9%, also described by OpenAI as saturating the test.
• ExploitBench: 100%, a cybersecurity-focused evaluation.

None of these figures have been independently verified. OpenAI has not published a technical report, methodology, or third-party evaluation alongside the announcement.

OpenAI also claims Astra outperforms Anthropic's Claude Opus 5 and an unspecified "Fable 5" model on computer-use tasks, while costing less to run. The company demonstrated Astra navigating Google Maps to locate a pediatrician, visiting the provider's website, and autonomously filling out a contact form with a care request message on a user's behalf.

OpenAI says Astra is also improved at coding, browsing, cybersecurity, and what it calls "the practical work behind scientific discovery," alongside better output quality for documents and presentations. The company frames this as enabling users to "delegate tasks with greater confidence in Astra's judgment," though no independent data on task-completion reliability or error rates was provided.

What's confirmed vs. claimed

Confirmed: OpenAI has announced GPT-6 Astra exists and is being distributed to a limited group of organizations as of September 3, 2026, with wider ChatGPT tier access to follow.

Unverified (OpenAI claims only): all benchmark scores (FrontierMath, ARC-AGI-3, ExploitBench), comparative performance against Claude Opus 5 and Fable 5, cost advantages on computer-use tasks, and claims of contributing to solved mathematical problems.

What this means

A 98% score on FrontierMath Tier 4 and 99.9% on ARC-AGI-3, if independently confirmed, would represent results well beyond what any publicly benchmarked model has achieved to date on these specific evaluations — both are designed to resist saturation by frontier systems. That gap between claim and independent verification matters: OpenAI released no technical paper, no red-team summary, and no third-party benchmark confirmation alongside the launch.

The staggered rollout — limited organizations first, broader ChatGPT access "in the coming days" — is standard practice for OpenAI's largest releases and likely reflects capacity constraints or continued safety testing rather than product readiness alone. Buyers and researchers should treat the benchmark figures as marketing claims until OpenAI publishes a model card or system card with methodology, and until outside groups such as METR, Epoch AI, or academic labs run comparable evaluations. Pricing and context window details, both critical for enterprise adoption, remain undisclosed.