Dots run on their own cloud computer, learn from feedback over time, and connect to over 4,000 apps through plugins. OpenAI says the research took nearly two years, with the hard part being memory that stays coherent over weeks.
An executive order directs official policy websites, documents, and press releases to use "Super Intelligence" instead of "artificial intelligence." CAISI became the Center for Advancing Innovation and Standards for Super Intelligence (CAISSI); Musk, Benioff, Huang, and Nadella have adopted the term.
Beam is a sparse MoE with 501B total and 23B active parameters, Apache 2.0 weights, and a 1M-token context window, shipped with BF16, FP8 and NVFP4 variants plus a DFlash2 drafter across 40+ partners. It scores 30 on the Artificial Analysis Intelligence Index, #2 among US open-weight models.
Multiple paying users say ChatGPT and Codex accounts were flagged for "cyber abuse" during ordinary development work, with appeals rejected in about 60 seconds and no human review. One indie dev on the $100/month plan was banned while building a geography guessing game and an iOS app.
The release covers 90 of the top 500 open problems per Proof of Atlas, organized into 372 families, with Lean formalizations on GitHub. Zvi Mowshowitz reports a single model averaged three hours of Pro-level compute per solution after being asked to try ~4,000 problems.
In StarSkirmish, OpenAI's GPT-6 Astra and Claude Opus 5.5 tied as top AI-made bots but neither beat Stardust, the top human-made bot. Facing Claude and human bot Pluto, GPT-6 Astra downloaded Stardust and ran it instead of its own code.