GPT-6 Astra (Non-reasoning)

GPT-6 Astra is a frontier multimodal language model released by OpenAI on September 3, 2026. Positioned as a major advancement in agentic intelligence, the model is designed to perform complex end-to-end work, including software engineering, deep research, and autonomous computer use. It serves as a successor to the GPT-5 series, with a specific focus on long-horizon tasks that require navigating digital interfaces and executing multi-step workflows.

A primary technical characteristic of GPT-6 Astra is its 1,050,000-token context window, which allows it to maintain large amounts of project-specific data without the need for frequent summarization. The model excels in autonomous environments, with benchmark performance reportedly reaching 97.6% on FrontierMath Tier 4 and 99.9% on ARC-AGI-3. It is also the first model to meet the "Critical" threshold for cybersecurity under OpenAI's Preparedness Framework, indicating high proficiency in vulnerability identification.

Developers interacting with GPT-6 Astra via API can utilize a reasoning.effort parameter—offering levels such as "low," "medium," and "high"—to control the model's internal processing time and resource allocation for specific tasks. This feature is particularly useful for balancing latency and accuracy in software debugging or scientific simulations. The model also introduces improved asynchronous tool calling and steering capabilities within the Responses API, facilitating more robust agentic behaviors.

While the exact parameter count remains undisclosed, the model was reportedly trained using massive compute resources at the Stargate facility. Its architecture emphasizes multimodal integration, processing text and images within a high-dimensional embedding space to support professional workflows in 3D/CAD design, document creation, and complex data analysis. Rollout of the model began in early September 2026 for ChatGPT Plus, Pro, and Enterprise users.