Morning Edition · Sunday, October 25, 2026
Tech
Moonshot's Kimi K3.1 Pro comes within a point of Claude Opus 5.5 on Terminal-Bench 4.0, with open weights due by November 4
The 2.8-trillion-parameter mixture-of-experts model scores 65.7% against Opus 5.5's 66.4% and GPT-6 Astra's 58.2%. A separate harness study shows orchestration alone can raise a 27-billion-parameter open model above a closed frontier model on hard coding problems.
Markets
Anthropic's draft listing filing commits it to $518 billion of computing over a decade, most of it non-cancelable
About 80% of the obligations to Google, Amazon, Microsoft, Broadcom, xAI and AMD must be paid whether or not Anthropic uses the capacity, against reported revenue of $4.6 billion and a possible listing valuation above $2 trillion.
Markets
Meta's free cash flow turns negative for the first time since its 2012 listing as AI capital spending rises
Shares fell nearly 7% on October 22, and the cost of insuring Meta's debt against default rose sharply after the company raised the floor of its 2026 capital spending guidance to $135 billion.
Markets
Hong Kong-listed MiniMax and Zhipu have lost a combined $150 billion from their peaks as Chinese token prices fall
MiniMax now trades below its January listing price, having been about 650% above it in March, and Zhipu is down 81% from its June high, with share unlocks approaching in January.
Tech
Claude produced assembly-level crash-freedom proofs for 239 entry points across 123 OSS-Fuzz libraries
Anthropic and the startup Theorem turned existing fuzz harnesses into formal specifications, generated about 180 million lines of Rocq proof and found bugs that years of fuzzing missed, though the proofs cover only single-threaded static RISC-V binaries.
Tech
Claude designed 17 confirmed binders against four scorpion neurotoxins in a 64-hour, mostly autonomous run
The internal research model generated 318,484 designs using 12 protein tools and about 10,000 GPU hours, but Anthropic has not tested whether any binder neutralizes a toxin.
Tech
Anthropic opens tiered access to its strongest cyber models as Moonshot strips hacking data from its open release
Anthropic's expanded Cyber Verification Program gives vetted defenders, red teams and testers of critical systems access to Claude Mythos 5.1 with reduced blocking classifiers, while Moonshot removed cyber data from Kimi K3.1 Pro before publishing it.
Tech
OpenAI's in-house storage platform passes 500 petabytes and 70 million requests a second as the company plans for exabyte scale
Habitat, which serves ChatGPT and Codex in about 40 regions, grew more than tenfold in a year, and two engineers rewrote it from Python to Rust using OpenAI's own tools.
Tech
Tencent open-sources Hy-Nomas, a steerable video world model that runs on a single RTX 4090
The Apache 2.0 model generates 720×1280 video at 24 frames per second one segment at a time and accepts new text or camera instructions mid-generation, using a persistent memory cache (a KV cache) that keeps the generated world consistent.
Tech
Block's Dessa learns conversation from audio alone as Ollama adds local speech synthesis
Block trained the full-duplex model from scratch on recorded human conversation using a single GB200 NVL72 system, while Ollama now serves Gradium's Phonon voices more than 100 times faster than real time on Apple Silicon.
Tech
Berkeley Lab and the University of Pittsburgh put Meta's SAM 3 and DINOv3 into beamline analysis and assistive robotics
The national-lab pipeline cuts segmentation of a reconstructed 3D volume from roughly a month of expert work per time step to about 15 minutes, while Pittsburgh runs the models on a battery-powered wheelchair.
Tech
A resurfaced benchmark claims more capable language models rate human nature more negatively
The test adapts a 1960s psychology questionnaire to score models on six dimensions of human nature, but it relies on self-report items that have not been shown to predict model behavior.