OpenAI launches GPT-6 with a built-in agent runtime
The new model ships with a persistent execution layer that lets agents run long-horizon tasks across thousands of steps — and a price cut that puts pressure on the rest of the field.
OpenAI unveiled GPT-6 on Wednesday, the company's first frontier model to ship with what it calls an integrated agent runtime — a server-side execution environment designed to let the model run multi-step tasks across hours or days without losing state.
The launch, announced at a closed press event in San Francisco, marks the most significant architectural change since GPT-4 and lands in the middle of an escalating race with Anthropic, Google DeepMind, and a half-dozen Chinese labs.
What's actually new
The runtime — internally codenamed "Driver" — pairs the model with a sandboxed compute layer that persists files, browser sessions, and intermediate reasoning across thousands of tool calls. Developers can hand off a brief like "monitor this product launch for the next 48 hours and ping me when something breaks," and the model coordinates the work without re-prompting.
OpenAI says GPT-6 also matches or beats Claude 4.8 on its internal long-horizon evaluations, including a new benchmark called LIVE-72 that measures sustained task completion over three days.
"The unit of work is no longer a token. It's a task." — Sam Altman, at the launch event
Pricing puts pressure on competitors
The API is priced at $4 per million input tokens and $12 per million output tokens — roughly half the equivalent rate for GPT-5 at launch a year ago. Industry insiders say the cut is aimed squarely at Anthropic's Claude 4.8, which has eaten into OpenAI's enterprise revenue throughout the first half of 2026.
Anthropic declined to comment. Google said GPT-6 "did not change our roadmap" for Gemini 3, expected later this summer.
What's next
GPT-6 begins rolling out to ChatGPT Plus and Team users today, with API access opening to the wait-listed Tier 5 developers tomorrow morning. The agent runtime is gated behind a separate beta and requires manual approval — a clear signal that even OpenAI is still nervous about what happens when a frontier model runs unsupervised for three days straight.
