📰 News
Claude Sonnet 5 — near-Opus performance at Sonnet prices
Anthropic shipped Claude Sonnet 5, its most agentic model yet at the Sonnet tier. It autonomously plans tasks, uses browsers and terminals, and hits 63.2% on SWE-bench Pro (Opus 4.8 scores 69.2%), while actually edging past Opus on the GDPval-AA v2 knowledge work benchmark. Introductory pricing: $2/$10 per million tokens until August 31. The catch: Sonnet 5 ships with a new tokenizer that uses ~30–40% more tokens on the same English text than Sonnet 4.6 (measured by Simon Willison). Nominal price looks the same; real cost is higher. — The Decoder
Claude Science beta — an AI workbench built for researchers
Anthropic is explicit: this is not a new model. The bet is on workflow integration — connecting 60+ scientific databases into one interface, bundling pre-built toolkits for genomics, protein structure, and cheminformatics, and running a dedicated verification agent to catch citation and calculation errors before anything gets published. Data stays local or on your HPC cluster. Up to 50 research projects can apply for $30K in credits (deadline: July 15). — TechCrunch
Google launches Nano Banana 2 Lite + Gemini Omni Flash API
Images in 4 seconds at $0.034 each; video at $0.10 per second of output. Google’s recommended pipeline: generate a still with Nano Banana 2 Lite, then animate it into video with Gemini Omni Flash. — The Decoder
Amazon AWS launches $1B forward-deployed engineering org
AWS is standing up a Palantir-style FDE unit: engineers embed inside client companies to build AI systems and leave behind lasting capability. OpenAI ($4B) and Anthropic ($1.5B) launched their own FDE ventures earlier — AWS is the third big player to follow. — TechCrunch
Meituan’s LongCat-2.0: 1.6T parameters, zero Nvidia
Chinese food-delivery giant Meituan trained a 1.6 trillion parameter model on a cluster of 50,000+ domestic AI chips, covering 35 trillion tokens. It beats Gemini 3.1 Pro and GPT-5.5 on SWE-bench Pro (59.5). It’s the most credible public demonstration yet that Chinese AI development can continue despite US export controls — though the model hasn’t been posted to HuggingFace, so independent verification is still pending. — The Decoder
SF rents: $180K salaries can’t find apartments under $5K/month
Average SF rent: $3,827. Median home price: $1.7M. Vacancy rates in hot neighborhoods dropped from 13% in 2020 to roughly 3% today. Pending IPOs for OpenAI and Anthropic — each valued near $1 trillion — could push prices even higher. — The Decoder
🔥 GitHub Trending
- agency-agents — A curated collection of specialized AI agent personalities that install into Claude Code, Cursor, Codex, and Gemini CLI in one click. Desktop app now available.
- FluidVoice — Offline voice dictation for macOS, fully on-device with a local AI model. No cloud, no API keys. Homebrew-installable.
- openpilot — comma.ai’s open-source driver assistance OS supporting 300+ cars. Repositioning as a general robotics operating system.
- CuPy — Drop-in NumPy/SciPy replacement that runs on NVIDIA CUDA or AMD ROCm GPUs, handy for GPU-accelerated ML pipelines.
▶️ YouTube
- “Telling my chickens Dario won” (Matthew Berman) — A lighthearted short celebrating the Claude Sonnet 5 launch. The chickens remain unmoved.
💬 Community
- What’s new in Claude Sonnet 5 — Simon Willison benchmarked the new tokenizer and found English text costs ~1.4x more tokens than Sonnet 4.6. Nominal price is unchanged; effective cost is not.
- The AI Compass — A political-compass-style quiz with 29 questions that classifies you into one of 30 AI ethics archetypes. Simon landed on “The Garage Tinkerer.”
- shot-scraper 1.10: have your agent record demo videos — New
shot-scraper videocommand takes a YAML storyboard and records a Playwright video of an agent walking through your app. Handy for automating demo production. - MAX models now run on Apple Silicon GPUs — Modular’s MAX runtime gains GPU acceleration on Apple M-series chips. Good news for local LLM runners on Mac.