The storefront for OpenAI’s newest flagship model gained some weight over the weekend. The company updated its product page for GPT-6 Astra with benchmark tables, descriptions of enterprise and safety features, and the pricing that developers will pay to build on the model. The refresh reads like a launch in progress: Astra has existed since Sept. 3, and OpenAI is now assembling the full commercial apparatus around it, from sales collateral to technical specifications.
The numbers on the page give outsiders their clearest look yet at where Astra sits relative to everything else on the market. On ARC-AGI-3, a reasoning test designed to resist memorization, OpenAI reports a score of 99.9 percent, close to the practical ceiling of the benchmark. On FrontierMath Tier 4, a collection of research-level mathematics problems, the company puts Astra at roughly 97.6 percent, a level that would have been hard to credit a year ago.
The comparisons matter more than the raw scores, because they are the terms on which OpenAI is asking customers to switch. On Terminal-Bench 4.0, which measures how well models operate real computer terminals to complete tasks, Astra scores 57.9 percent, against 37.3 percent for OpenAI’s own previous flagship, GPT-5.6 Sol, and 55.8 percent for Anthropic’s Claude Fable 5.1. On the Agents’ Last Exam, a professional-task benchmark, Astra’s 59.3 percent edges out the 55.5 percent reported for Claude Opus 5.
The spread on Terminal-Bench is the number that agents builders will study longest. A twenty-point gap over Sol suggests the model’s ability to operate software has improved sharply in a single generation, and the narrower lead over Claude Fable 5.1 frames the race with Anthropic as close at the frontier. OpenAI published both comparisons on the same page, an implicit claim that Astra is now the model to beat on the kind of work that runs a computer rather than merely answering questions.
The page also fills in the commercial picture. OpenAI has set API prices of $10 per million input tokens and $50 per million output tokens, above the rates its predecessor commanded, and it is describing the model’s enterprise and safety capabilities for the procurement officers and security teams who will decide whether their companies adopt it. The company has acknowledged that Astra’s reasoning is harder for outsiders to audit, and the safety section of the page is where it makes its case that the trade is worth making.
Rollout plans published alongside the benchmarks show a deliberate staircase. Astra will open in the coming days to Plus, Pro, Business and Enterprise subscribers, the tiers that pay OpenAI directly, before reaching developers through the API and enterprises through Microsoft’s Azure and Amazon’s AWS Bedrock. OpenAI has not said when every channel will be live. The staged pattern suggests the company is rationing access as it brings capacity online, a familiar discipline for a lab whose newest models tend to arrive in waves.
Timing is doing some of the marketing work. The page refresh lands less than a week after Astra’s debut and days after OpenAI disclosed the price ladder for the model. In the competitive rhythm of the AI industry, where model releases are followed within days by rival releases and price cuts, a complete storefront matters as much as the research behind it: customers decide what to build on the information in front of them.
The enterprise emphasis reflects where the money is. OpenAI’s growth has come increasingly from businesses and developers who embed its models in products, and those buyers ask different questions than consumers do. They want to know how a model behaves under a security audit, whether it can be deployed in a regulated industry, and what the economics look like at scale. A product page is the first place those answers appear, and OpenAI is using Astra’s page to preempt the questions its sales force will hear.
Benchmark pages invite their own skepticism, and analysts who follow model evaluation caution that the tests are only a partial map of capability. ARC-AGI-3 rewards certain styles of reasoning, FrontierMath samples a narrow slice of mathematics, and terminal tasks measure one mode of agency. Still, the direction of the numbers is hard to argue with: Astra clears its predecessors on the tests OpenAI chose to publish, and where rival scores are known, it generally clears them too.
What the page does not yet contain is a date for full availability. OpenAI has said Astra will reach every channel in the coming period, without committing to a calendar. For a model this large, the constraint is usually compute: inference capacity has to be stood up before a product can be opened to everyone. The staged rollout, the published benchmarks and the pricing table together form a familiar pattern, a company preparing its biggest product for general release while managing the infrastructure that makes that possible. The spec sheet is now public. The queue is not.
