Claude Opus 4.8, NVIDIA's Cosmos 3 keynote, Codex on Windows

What shipped

Anthropic

Claude Opus 4.8

A May 28 upgrade priced like Opus 4.7, with effort defaulting to high and about four times fewer unflagged flaws in its own code, per Anthropic.

NVIDIA

Cosmos 3

An open omni-model that combines world generation, physical reasoning and action in one mixture-of-transformers, announced at Sunday's GTC Taipei keynote.

Mistral AI

Introducing physics AI at Mistral

Mistral brings in Emmi AI to build models that predict full physical fields in seconds on one GPU, with ASML, Airbus, Safran and Siemens Energy named as partners.

StepFun

Step 3.7 Flash

StepFun's May 28 release, with a model page on Hugging Face.

OpenAI

Computer use and remote control for Codex on Windows

The May 29 update brings computer use and remote control to Windows.

arXiv

HRM-Text: Efficient Pretraining Beyond Scaling

A pretraining approach pitched on efficiency instead of raw scale, first on the weekly papers list.

arXiv

Gamma-World: Generative Multi-Agent World Modeling Beyond Two Players

A generative world model built for more than two interacting agents.

Opus 4.8 landed on Thursday, and the week ended with a keynote on a Sunday night.

Opus 4.8

Anthropic released Claude Opus 4.8 on May 28 at the same $5 and $25 per million tokens as 4.7. The launch post describes it as a modest but tangible improvement and says it makes about four times fewer unremarked flaws in its own code than Opus 4.7. It defaults to high effort, adds extra and max levels, and accepts system messages inside the conversation so instructions can change mid-task. The API notes add a 1M-token context by default, images up to 2,576 pixels on the long edge, and a research preview of fast mode.

For Claude Code, Anthropic expanded Auto mode, set Max users to fast mode on Opus 4.8 by default, and opened Workflows in research preview for very large tasks. One reported number: 84% on Online-Mind2Web for computer use. I would read that as a vendor chart, but the direction fits what I see when an agent has to fill in a supplier portal.

NVIDIA's Sunday keynote

NVIDIA's GTC Taipei keynote ran Sunday evening Pacific time and announced Cosmos 3, an omnimodel built on a mixture-of-transformers design that combines world generation, physical reasoning and action. NVIDIA put the model on Hugging Face the same day with an open toolkit for data generation, simulation, training and validation. The keynote also introduced an open humanoid reference design on Jetson Thor. I will have more on the rest next week.

Smaller items

Mistral announced physics AI on May 27 after bringing in Emmi AI. The models learn from solver outputs and predict full physical fields from geometry in seconds on a single GPU, against hours or weeks for a conventional CFD or finite-element run. ASML, Airbus, Safran and Siemens Energy are named. That is a serious industrial claim and I have no way to test it.

Codex reached Windows for computer use and remote control on May 29. StepFun released Step 3.7 Flash on May 28, with a model page on Hugging Face. HRM-Text led the weekly papers.

What I'd watch next: the rest of NVIDIA's Computex package, especially which models ship with weights the same week.