OpenAI Astra

OpenAI Astra is the company's next major model. OpenAI has published mathematics results from an internal version and designated Astra as its first model to meet the Critical cybersecurity capability threshold. As of September 3, 2026, OpenAI says it plans to make Astra available soon; a public API model ID, pricing, context window, and final System Card have not been published.

OpenAI calls Astra its next major model. The company attributes ten mathematics and theoretical-computer-science results to an internal Astra version and has published manuscripts plus Lean certificates. OpenAI later said Astra meets the Critical cybersecurity capability threshold under its Preparedness Framework.

OpenAI reports 100% on ExploitBench and stronger performance than GPT-5.6 Sol on an internal set of 20 recent high-severity V8 vulnerabilities. The company says Astra found two zero-day vulnerabilities during evaluation. These are provider-reported results, and OpenAI notes that highlighted results used Daybreak Blue access rather than the default production configuration.

OpenAI plans restricted access for advanced cyber workflows, beginning with selected alpha testers and expanding defensive access through Daybreak Blue. General deployments are expected to include stronger refusals, action monitoring, and human-review controls.

OpenAI has not confirmed GPT-6 as Astra's shipping name. Astra is the official entity name currently used by OpenAI; GPT-6 remains a search and reporting label until release documentation establishes the final family name.

Reporting has linked Astra to recurrent depth, a technique that reuses computational blocks to spend more latent computation on difficult inputs. Public research shows this architecture is plausible, but OpenAI's Astra disclosures do not confirm the implementation. See the linked Atoms analysis for the evidence and limitations.

Teams should verify the exact model ID, price, context and output limits, tool interfaces, regional access, latency, safeguard behavior, and rollback strategy before production migration. Keep a supported current model as a fallback until Astra passes workload-specific offline, shadow, and canary evaluations.