OpenAI releases GPT-6 Astra, its most capable model, in general availability

OpenAI has begun shipping GPT-6 Astra, the model it describes as "our most intelligent model yet, with state-of-the-art performance in computer use, browsing, software engineering, science, and professional work." The rollout started September 3, 2026, with the OpenAI API changelog confirming the release and enterprise access opening first through the company's Trusted Access Program.

What's new

GPT-6 Astra is positioned as an end-to-end task model rather than a chat-first incremental upgrade. Per OpenAI's own developer documentation, it is "our most capable model, built for the hardest end-to-end work," intended for "reasoning, coding, computer use, research, and document creation," combining those capabilities "to carry complex tasks from an initial request to a finished result, using the context and tools you provide."

Notable specifics from the changelog and model guide:

• Async tool calling and mid-turn steering, aimed at long-running agentic workflows.
• Structured outputs and prompt caching support.
• OpenAI says the model "achieves stronger results while using substantially fewer output tokens — delivering a lower estimated API cost per task than earlier models despite its higher per-token pricing."
• Several behavior changes ship alongside the model: it does not support the none reasoning effort level, does not support custom temperature or top_p values or log probabilities, and tool calling now requires the Responses API rather than Chat Completions.
• A new "misalignment monitoring" system runs asynchronously during agent work on supported Responses API requests, and can trigger safety alerts or halt a conversation for review.

Availability is staged: enterprise customers in the Trusted Access Program get it first, with API and subscription-plan access (Plus, Pro, Business, Enterprise) following "in the coming days." Developers can already target it by setting model to gpt-6-astra in API requests.

Context

The release follows OpenAI's September 1 disclosure that Astra's system card marked it as the first OpenAI model to cross the company's "Critical" cybersecurity capability threshold — a safety classification, not a launch. This week's changelog entry and model guide mark the actual shipping of the model into developers' and enterprises' hands, with OpenAI simultaneously emphasizing new guardrails (the misalignment monitor, tightened API constraints) alongside the capability jump.

OpenAI frames the alignment work as inseparable from the capability increase, stating plainly that "GPT-6 Astra is also our most aligned model yet. It excels at exercising care, respecting task boundaries, and communicating transparently."

Why it matters

Shipping a model that OpenAI itself has classified as crossing a "Critical" capability threshold is a significant test of how the company operationalizes its own safety commitments in production, not just in a system card. The staged rollout — enterprise-first, broader API and consumer access later — suggests OpenAI is treating deployment itself as part of the safety process rather than a single release event. For developers, the API-level changes (no Chat Completions tool calling, no custom temperature/top_p, mandatory migration considerations) mean GPT-6 Astra is not a drop-in replacement; teams building agentic or long-horizon workflows will need to evaluate the Responses API migration path before adopting it.