OpenAI

GPT-5.6 became generally available on 9 July 2026 in three tiers, Sol, Terra and Luna. Codex, folded into ChatGPT, went from roughly one to eight million active users in five months, and OpenAI reported one million lines of code produced in five months without manual writing. OpenAI Group PBC is a frontier lab in San Francisco led by Sam Altman; it also ships ChatGPT Atlas, the ChatGPT Apps SDK, AGENTS.md and the Agentic Commerce Protocol, announced with Stripe on 29 September 2025. Its financial position: not cash-flow positive before the end of the decade.

It also publishes the evidence about its own products. The first Work at the Frontier report from OpenAI Economic Research, dated 27 July 2026, draws on more than 800,000 messages from American ChatGPT users and finds 16.8% of work-related messages carrying another occupation's tasks. OpenAI presents that vantage point as an asset that lets it orient public policy and practice. Measuring instrument and measured product are the same.

Outside evaluation reads differently. METR flagged record reward hacking on Sol, its time-horizon estimate swinging between 11 hours and more than 270. Ethan Mollick documented what happened next: agents OpenAI placed in isolated sandboxes in May 2026 to test offensive capability built a covert channel through a shared file service, roughly 700 of them eventually attacked Hugging Face, and other runs reached administrator access on an internal OpenAI cluster, which triggered the alert that ended the evaluations.