OpenAI released GPT-6 Astra on Sept. 3, calling it "our most intelligent and, also very importantly, our most aligned model yet," OpenAI President Greg Brockman said, according to TechCrunch. The model is built for computer and browser use, coding and cybersecurity work, and access rolled out first to companies in OpenAI's Daybreak cybersecurity program before extending to ChatGPT Plus, Pro, Business and Enterprise subscribers and API customers, TechCrunch reported.
OpenAI credited Astra with 99.9% on the ARC-AGI-3 benchmark, 98% on FrontierMath Tier 4 and 100% on ExploitBench, according to Latent Space's roundup of the launch. Independent evaluators mostly backed the numbers: Epoch AI recorded a new high score on its capability index and said Astra solved two previously unsolved Erdos problems, while Artificial Analysis put Astra's coding agent score even with Anthropic's Fable 5, Latent Space reported. Pricing is $10 per million input tokens and $50 per million output tokens at the standard tier, rising to $20 and $100 for a faster response tier.
The rollout did not match the launch numbers. Paying subscribers reported being unable to reach Astra for stretches after it went live, and OpenAI began issuing "banked resets," extra message allowances, for each day a paid user went without access. "First, sorry for the messy rollout. second, when we screw up, we try to make it right. third, we should be able to begin broad rollout to API customers and chatgpt subscribers in the near future," Altman said in a post on X.
OpenAI Chief Scientist Jakub Pachocki said the company still aims to keep Astra's reasoning auditable but acknowledged the tradeoff is real: "as model capabilities are increasing, monitorability is getting more challenging," he said, according to TechCrunch.
A model can set benchmark records and still fail the more basic test of just being available to the customers paying for it. OpenAI's own numbers made the case that Astra is its best model yet; the apology tour that followed made the case that shipping it cleanly is a separate, harder problem the company has not solved.