GPT-6 Astra Launch: Price, Benchmarks, Access

OpenAI has released its next flagship frontier model. On September 3, 2026, the company announced GPT-6 Astra, describing it as its most intelligent and capable system to date and framing the launch in unusually sweeping terms — president Greg Brockman called it a “generational leap” and said it marks the start of the “AGI era.” The model began rolling out the same day to a limited set of organizations, with broader availability across ChatGPT and the API promised in the days that follow, according to reporting from Bloomberg, Axios, and 9to5Google.

Astra is also a first of a different kind. OpenAI says it is the first of its models to reach the “Critical” cybersecurity threshold under its Preparedness Framework, the internal risk classification that determines when a capability is dangerous enough to require extra safeguards before — or instead of — a normal launch. That designation shapes how the model ships: its most advanced cyber capabilities are gated rather than turned on for everyone.

What OpenAI announced

GPT-6 Astra succeeds the GPT-5.6 “Sol” line and is positioned as a step change in reasoning, agentic work, and autonomous tool use rather than an incremental point release. OpenAI is pitching it as a system that can carry out long, multi-step tasks with less human supervision — planning, writing and editing code, operating desktop applications, and working through research-grade problems end to end.

The rollout is staged. Access opened first to a limited set of organizations and vetted partners, with OpenAI saying general availability for ChatGPT Plus, Pro, Business, and Enterprise users and API developers would arrive in the “coming days,” alongside a wider public release. That cautious sequencing is deliberate, and it ties directly to the safety classification described below.

On price, GPT-6 Astra runs $10 per million input tokens and $50 per million output tokens through the API, with a context window of roughly 1.05 million tokens. A “fast” processing mode runs at up to 2.5x the speed of standard for 2x the price. Independent trackers noted the model is meaningfully more expensive than Sol at maximum effort — a jump driven by a roughly 2.5x increase in per-token price, partly offset by Astra using fewer tokens to reach an answer. Notably, the new pricing matches Anthropic’s rate for Claude Fable 5.1, putting the two leading labs’ flagship API prices in the same tier.

The benchmark claims

OpenAI and early independent evaluations point to strong results on the tasks the industry now treats as frontier tests — computer use, agentic coding, and hard mathematics.

On OSWorld 2.0, a computer-use benchmark that measures how well a model can operate a real desktop environment, Astra is reported to score 72.6% while taking roughly 47% less time per task than its predecessor. On DeepSWE v1.1, a 113-task agentic coding benchmark, it scores 74.1%. In mathematics, OpenAI says the model effectively saturates FrontierMath Tier 4 at 97.6% — the hardest tier of a benchmark built specifically to resist memorization — and posts a perfect 100% on ExploitBench, a security-focused evaluation.

That last figure is the tell. A model that tops a benchmark for finding and exploiting software flaws is exactly the kind of system OpenAI’s safety framework was built to flag. The company has been laying groundwork for this moment for months: it detailed Astra’s ability to solve long-open math problems with machine-checkable Lean proofs in early August, and evaluators have watched the model’s autonomous problem-solving climb since.

Why the “Critical” label matters

The most consequential part of the launch is not a benchmark — it is the classification. OpenAI says Astra is the first model to hit the Critical cybersecurity tier under its Preparedness Framework, meaning it can, under the right conditions, find previously unknown vulnerabilities and build working exploit chains against well-defended systems without a human directing each step.

That capability is dual-use by nature: the same skill that lets a red team surface zero-days before adversaries do also lowers the barrier for those adversaries. OpenAI’s response is controlled distribution rather than suppression. The model’s most advanced offensive and defensive security features are being withheld from the general consumer version and routed instead through Daybreak Blue, the approved-tester track OpenAI uses to put its strongest cyber tooling in the hands of vetted defenders.

None of this is improvised. OpenAI previewed the restrictions on September 1, and weeks earlier it had paused a frontier reinforcement-learning run as evaluations approached the Critical line. The September launch is the resolution of that pause: rather than hold the model indefinitely, OpenAI is shipping it with the dangerous capabilities gated and additional production safeguards — the company cites alignment training, automated code review, and misalignment monitoring — layered on top.

The AGI framing

Alongside the technical detail, OpenAI made a rhetorical bet. Brockman’s “AGI era” language, and reporting that the company believes Astra “may represent” a form of artificial general intelligence, are the strongest such claims OpenAI has attached to a shipped product. The framing is contestable — there is no agreed definition of AGI, and independent analyses noted that Astra sits behind Sol on some cost-adjusted intelligence measures once its higher price is factored in. A model can be the most capable ever released and still not be “general” in the sense skeptics demand.

What is not in dispute is the trajectory. Each flagship release from the leading labs now lands with longer task horizons, higher benchmark ceilings, and — increasingly — a governance story attached. Astra arrives days after Anthropic shipped Claude Fable 5.1 and Mythos 5.1 and after Google’s Gemini 3.8 Flash release, keeping the three-way frontier race at a pace where a new “most capable model” is measured in weeks, not quarters.

What it means

GPT-6 Astra’s launch is two stories in one: a capability milestone and a governance test.

For enterprises and developers, the value is in agentic reliability, not raw novelty. The figures that matter here are the computer-use and coding scores paired with the reduced task time — signals that the model can be handed longer, messier jobs and finish more of them without a human babysitting each step. That is what justifies a higher token price to a buyer running agents at scale: fewer failed runs and less human cleanup can outweigh a steeper sticker cost. Teams evaluating Astra should benchmark it on their own workflows against Sol and Anthropic’s flagship rather than trusting headline numbers, because the cost-per-task math is workload-specific.

For the safety debate, Astra is the first real stress test of self-imposed limits. Declaring a model “Critical” and then genuinely restricting its distribution is exactly the constraint OpenAI’s framework was supposed to produce. Whether the company holds that line under commercial pressure — keeping the rawest cyber capability behind Daybreak vetting rather than quietly loosening it — is what will determine if the Preparedness Framework reads as real governance or a launch-day talking point. Critics who see safety framing as cover for keeping powerful models proprietary will be watching the enforcement, not the announcement.

For the competitive picture, the pricing convergence is the quiet headline. With Astra matching Anthropic’s Fable tier, the two leading labs are no longer competing much on flagship API price — they are competing on capability per dollar and on trust. Expect the next round of differentiation to come from agentic dependability, integration depth, and how each lab handles the dual-use capabilities that Astra just made unavoidable.

The things to watch next are timing and scope: how quickly Astra reaches general availability, how wide the Daybreak Blue circle grows, and whether independent evaluators confirm the benchmark claims once the model is broadly in the wild. A system that can find and weaponize zero-days without a human at the keyboard does not stay contained by policy language alone — it stays contained by how rigorously that policy is enforced.