What's New in AI

Cloudflare released Clef and Clef-flash, open-weight decision models on Workers AI, with vendor-reported wins over Jev and a price that drew immediate pushback.

Figure AI retired Figure 02 by teaching the robots to jump into molten steel in Finland, then sold the metal. The film is real; the reason is ordinary.

Pi 1.0 ships MCP support by default, reversing the minimal coding agent's own position. We read the changelog, the reversal post, and the skeptic case.

OpenAI's GPT-6 Astra proved prime gaps at most 186: infinitely many pairs of consecutive primes differ by 186 or less. What that means, and what it does not.

Tavus says Griffin is the first model to pass a real-time video Turing test, with 48% of its own study's participants fooled. We separate claim from evidence.

Claude claims a proof of the percolation dying-conjecture, backed by a Lean file. We separate the machine check that compiled from the review still underway.

Artificial Analysis measured Gemini 4 Argon's hallucination rate at 15%, the lowest among top-tier models. But the same run shows its accuracy fell to 50%.

Google launched Gemini 4 Argon with a 1M-token output limit and top-tier scores. We separate the vendor benchmarks from independent tests and the skeptic case.

AMD has announced nothing yet about RDNA 5. We separate a real architecture from the spec list the leak mill circulates, and the arithmetic underneath it.

A GeForce RTX 6090 leak points to 192 SMs, a 512-bit bus and 32 GB of GDDR7 on Rubin. Nvidia confirms nothing, and the spec sheet is smaller than it sounds.