Agent & Model Releases
Google launches Nano Banana 2 Lite, its fastest and most cost-efficient image model, and makes Gemini Omni Flash available to developers for video generation and editing.
Anthropic releases Claude Sonnet 5, with performance close to Opus 4.8 at lower prices, and lifts export controls on Claude Fable 5 and Mythos 5.
Mistral AI open-sources Leanstral 1.5, a Lean 4 code agent for automated theorem proving, with 119B parameters and Apache 2.0 license.
Anthropic launches Claude Science beta, a multi-agent workbench for genomics, proteomics, and cheminformatics with over 60 curated skills and audit trails.
WebBrain is a free open-source browser agent for Chrome and Firefox that can run locally, automate tasks, and operate inside authenticated sessions.
Tools & Frameworks
LlamaIndex publishes legal-kb, a reference app for agentic retrieval using filesystem-style tools (grep, find, read) over a legal knowledge base.
sqlite-utils 4.0rc2 is mostly written by Claude Fable for about $149.25, catching critical bugs like missing commits that could cause data loss.
shot-scraper 1.10 adds video recording of agent demos via storyboard config files, enabling agents to produce visual walkthroughs of their work.
A tutorial demonstrates schema-guided invoice extraction with lift-pdf, handling tricky fields like bill-to vs. ship-to and partial payments.
llm-coding-agent 0.1a0 is a new coding agent built on the LLM library, supporting file editing, command execution, and a YOLO mode.
Research Highlights
EvoPolicyGym benchmarks how agents iteratively improve policies in interactive RL environments; GPT-5.5 achieves top performance on all 16 tasks.
MrFlow accelerates text-to-image flow models (FLUX.1-dev, Qwen-Image) by 5x+ without training using staged low-to-high-resolution sampling.
AgenticDataBench provides a comprehensive benchmark for data science agents across 15 domains, including real fintech B2B use cases.
ELDR improves MoE model serving by routing requests based on expert locality, reducing median token-by-token latency by up to 14%.
WorldDirector decouples semantic motion from video generation, enabling persistent dynamic object memory and controllable viewpoint exploration.
Industry & Business
Google DeepMind partners with A24 for a first-of-its-kind research collaboration to develop AI tools shaped by filmmakers.
Meta launches Pocket, a vibe-coded gaming app that lets users generate interactive games with AI prompts, built on Gizmo acquisition.
Microsoft launches Frontier Company with $2.5B commitment and 6,000 experts for enterprise AI deployments, similar to forward-deployed engineering models.
Venice AI becomes a unicorn with $65M Series A, offering uncensored open models and privacy-preserving access to 200+ AI models.
Alibaba bans employees from using Claude Code starting July 10, pushing its internal Qoder tool instead.
Cloudflare will block mixed-use crawlers (combining search, training, and agent use) from ad-supported sites by default starting September 15.
Opinions & Analysis
Former Qwen lead Junyang Lin shifts focus from training models to training agents, arguing hybrid thinking modes are not enough.
Better models can be worse at using specific tool schemas: newer Claude models invent extra fields in Pi's edit tool, unlike older ones.
Course creator Josh Comeau reports 50%+ revenue drop as learners switch to LLMs for tutoring and worry about developer job prospects.
Claude Code tip: let Fable use its own judgement to choose lower-power models for subtasks, saving tokens and improving efficiency.
Meta CEO Mark Zuckerberg tells staff AI agent development hasn't accelerated as expected, and job cuts weren't as clean as hoped.
That's everything for this day – about a 5-minute read.