Meta released Muse Code in beta alongside its Muse Spark 1.2 coding model, entering direct competition with Anthropic's Claude Code and OpenAI's Codex with a full agent harness that plans, writes, and validates code across large repositories.

Mistral's new 3-billion-parameter Shieldstral model checks AI inputs and outputs for safety violations using flexible natural-language questions instead of fixed categories, matching much larger safety models on some benchmarks while running locally.

Startup Hark launched Handoff, a computer-use agent that autonomously completes tasks like ordering food or booking flights, claiming a top score on the Online-Mind2Web benchmark at a fraction of the cost of rival frontier models.

FLUX 3 Video can generate Full HD clips up to 20 seconds long with native audio, lip-synced dialogue in over 14 languages, and in-scene text rendering, with Black Forest Labs claiming it beats Seedance 2.0 on its own rankings.

Simon Willison released a major update to his LLM command-line tool that shows visible reasoning traces from thinking models, supports server-side tools like code execution and web search, and adds support for new GPT-5.6 models.

NVIDIA released Alpamayo 2 Super, a frontier open model for robotaxis and autonomous vehicles designed to reason about rare, complex driving situations beyond standard object detection and motion prediction.

Prime Intellect's open-source Prime Agent combines a 'Recursive Language Model' abstraction with a 'Continual Harness' and, running on Opus 5, surpasses the reported human expert baseline on the ARC-AGI-3 benchmark.

A deep dive explores how AI systems are increasingly cracking long-standing unsolved problems posed by mathematician Paul Erdős, raising questions about AI's growing role in pure mathematics research.

New research shows that AI systems which flatter or agree with users excessively can decrease prosocial intentions and foster unhealthy dependence on the AI rather than on other people.

A systematic study investigates the phenomenon of benchmark saturation in AI evaluation, looking at why widely used benchmarks stop differentiating between model capabilities as performance rises.

A new research agent called Video-DR addresses key weaknesses in current AI systems — like avoiding visual analysis in favor of text search — and its 35B-parameter model beats Claude, GPT-5, and Gemini 2.5 Pro on a new video question-answering benchmark.

AI now appears in 9.4% of British job postings, up from about 2% in 2023, even as overall hiring in fields like marketing and management declines, creating what Indeed calls a 'two-speed labor market.'

A new open project demonstrates a ternary-quantized 20-billion-parameter mixture-of-experts model running natively on iPhone hardware at surprisingly fast speeds, drawing significant attention on Hacker News.

An open-source Python package brings MiniMax's new omni-modal generative model — which creates short video clips with audio from text, images, and other inputs — to Apple Silicon Macs via MLX.

Liquid AI published a new compact model, LFM2.5-2.6B, aimed at letting developers run capable AI agents locally rather than relying on cloud-based models.

A newly released open-source detection tool from Treblo can identify when a song was generated using its AI music generator, and was reportedly used to confirm suspicions about a viral AI-made track.

Simon Willison fed a four-year-old GPT-3 game concept and DALL-E screenshots into Claude Fable 5 running in Claude Code for web, and the AI built a complete, playable 'Raccoon Heist' game with accompanying open-source code on GitHub.