Anthropic

Anthropic opens a standard for AI agents to operate laboratory equipment

The Model Hardware Standard research preview covers programmable devices from microscopes to robotic arms. The company says integration drops from weeks to hours, and the standard is model-agnostic.

Meta

Meta's Muse Image lands on the API: it reasons before it renders, at a cent an image

Each call searches the web and refines the output over iterative passes, with extra evaluation of the prompt for precise elements like charts and QR codes. Generation and editing sit in one model.

Wan

Alibaba's Wan 3.0 goes generally available, turning documents into 30-second video

The model produces 30-second 1080p clips with audio in a single pass. Its distinguishing feature is Omni-Reference: documents, spreadsheets and slide decks count as inputs alongside text and images.

Apodex

Apodex 1.1: a model family that coordinates an agent team, with an open-weight small version

The flagship targets long-horizon professional work; the 35-billion-parameter Mini is released under Apache 2.0. FrontierAgent, the agent workbench, is open source too.

AMD

AMD targets its software gap with ROCm 10, putting its own skills into Claude and Cursor

The release adds a command-line tool, an AMD Skills pack that plugs into coding agents, and the open-source Hyperloom. AMD reports a 3.3x inference speedup over ROCm 7.0.

Cartesia

Cartesia's Sonic-3.6 tops both Artificial Analysis speech leaderboards

The company says it stopped fine-tuning the existing approach and rebuilt from the architecture up. The model leads both the Provider Voice and Controlled Voice arenas, but is in beta on Cartesia's API only.

alphaXiv

alphaXiv puts an army of agents to work reproducing arXiv papers, with Tinker paying the compute bill

The new autoresearch feature hands a paper to Claude and Codex agents and lets them run fresh experiments on it. Training runs go through Thinking Machines' Tinker API.

Goodfire

Goodfire: a handful of forking words decide where a reasoning chain lands

The company resamples every step of a model's chain of thought to find the tokens that actually change the outcome. It has also made its research agent Silico generally available.

Hark

Hark signs a multi-year deal with NVIDIA for gigawatt-scale compute on Vera Rubin

The new company from Archer and Figure founder Brett Adcock is building a multimodal assistant that remembers its user. The agreement covers both model training and the inference behind its launch.

Argmax

Argmax gets Qwen3-ASR running inside the iPhone, with speech transcribed on the device

Pro SDK 3 has reached a stable release. The multilingual recognition model runs on the phone rather than a server, and the company has opened a TestFlight app for trying it without writing code.

Google

Google Research publishes GlucoFM, a foundation model for continuous glucose data

Trained on 109,000 hours of unlabelled CGM data, the model splits the glucose curve into a slow baseline and short-term deviations. It is a research contribution, not offered as a product or clinical tool.

Anthropic

Fable 5.1 claims circulate, with no confirmation from Anthropic

Accounts following the AI industry claim the model is being tested with a limited group of users and that a launch is close. There is no company announcement, model card or pricing page.