GPT-6 Astra

• 🆕 OpenAI's flagship GPT-6 model, released September 2026.
• 📏 1.05M-token context: 922K input, 128K output.
• 🧠 Built for complex reasoning, coding, computer use, research, documents.
• 👁️ Accepts text and image inputs; supports function calling and web search.
• 🎯 On OpenAI's reported Terminal-Bench 4.0, 57.9% versus 37.3% for GPT-5.6 Sol.
• 💬 Asks clarifying questions more often; stays coherent on long tasks.
• 🔧 New Codex context preservation replaces summary-style compaction in long sessions.
• 🏢 Staged rollout: Trusted Access Program first, then API and paid plans.

OpenAI

OpenAI is an American artificial intelligence research organization headquartered in San Francisco, structured as both a for-profit public benefit corporation and a nonprofit foundation. The lab developed the GPT family of large language models, the DALL-E…

GPT-6 Astra opens OpenAI's GPT-6 generation and is positioned by OpenAI as its most capable model, aimed at "the hardest end-to-end work": complex reasoning, software engineering, computer use, research, and document creation. In the API it is presented as the default starting point for complex reasoning and coding tasks. It accepts text and image inputs, supports tool/function calling and web search, and exposes a 1,050,000-token context window with up to 128,000 output tokens.

Compared with the July 2026 GPT-5.6 line — GPT-5.6 Sol, GPT-5.6 Luna, and GPT-5.6 Terra, plus their Pro variants such as GPT-5.6 Sol Pro — OpenAI's own model guidance says Astra is more intelligent and capable than GPT-5.6 Sol and earlier models, stays more coherent during long tasks, follows longer instructions better, and is more likely to ask clarifying questions instead of assuming. On OpenAI's reported figures, Astra reaches 57.9% on Terminal-Bench 4.0 against 37.3% for GPT-5.6 Sol, and 100% versus 78.5% on the company's ExploitBench evaluation.

Architecturally, the notable operational change is how long sessions are handled: with Astra, Codex uses a new mechanism to preserve and retrieve context once the window fills, rather than the summarizing compaction earlier models relied on. OpenAI also describes stronger visual judgment for generated interfaces and the ability to build and host sites and apps from a prompt.

OpenAI rolled the model out first through its Trusted Access Program before broader API and paid-plan availability. Earlier generations such as GPT-5.5 Pro and GPT-5.3 Codex remain available for lighter or cheaper workloads.