Morning brief

Saturday, July 4

3 people · 0 companies · 4 topics

TL;DR Claude Sonnet 5 and Meituan’s LongCat-2.0 (1.6T params, Chinese chips, MIT license) both shipped June 30; Meta’s in-training Watermelon model claims GPT-5.5 parity and OpenAI is pitching a 5% US government equity stake.

Models & research

  • Anthropic launched Claude Sonnet 5 on June 30, now the default for Free/Pro plans and Claude Code, at $2/$10 per Mtok intro pricing through Aug 31 (then $3/$15). simonw flagged a practical catch: the same input generates ~30% more tokens than Sonnet 4.6, so per-task costs may exceed the headline price cut. model-releases Anthropic · TechCrunch · simonwillison.net

  • Meituan open-sourced LongCat-2.0 (June 30): 1.6T-param MoE with a native 1M-token context, MIT licensed, trained on 50k domestic Chinese ASICs and no Nvidia GPUs. It had been leading OpenRouter’s coding charts for two months under the alias “Owl Alpha.” SWE-Bench Pro score 59.5 beats GPT-5.5’s 58.6. Weights are “coming soon” on HuggingFace; hosted API is live. model-releases VentureBeat · SiliconAngle

  • elonmusk announced June 28 that Grok 4.5 (1.5T-param V9 + Cursor coding data) entered private beta at SpaceX and Tesla. Musk claims performance “close to, perhaps exceeding” Claude Opus. No public API, no independent benchmarks exist. model-releases TechTimes · LetsDatScience

  • Meta AI chief Alexandr Wang told staff July 2 that the in-training “Watermelon” model matched GPT-5.5 on internal benchmarks, using one order of magnitude more compute than the prior “Avocado” model. Caveat: OpenAI is already previewing GPT-5.6, so the target has moved. model-releases Benzinga · American Bazaar

Tools & dev

  • claude-code updated to v2.1.197: Sonnet 5 is now the default model, new Manual permission mode ships on by default, background subagents run without blocking the main session, and transient 429 rate-limit errors retry automatically with backoff. GitHub releases · Releasebot

  • Anthropic started blocking Chinese firms that used Singapore subsidiaries and VPNs to access Claude, following FT reporting that exposed Ant Financial and ByteDance doing this at scale. New policy: any company with 50%+ indirect Chinese ownership is banned globally; Anthropic is monitoring time zones and usage patterns to detect violations. BanklessTimes · Investing.com

OSS

  • usestrix/strix (1,910 stars today) — open-source AI pentest tool that autonomously finds and fixes app vulnerabilities. GitHub
  • alibaba/page-agent (1,110 stars today) — TypeScript in-page GUI agent for controlling web interfaces with natural language; designed for agentic browser workflows. agentic-ai GitHub
  • mattpocock/skills (1,289 stars today) — Matt Pocock’s Claude Code skills collection for engineering productivity. claude-code GitHub
  • JuliusBrussee/caveman (1,089 stars today) — Claude Code skill that cuts token usage ~65% through simplified communication protocols. claude-code GitHub
  • openai/codex-plugin-cc (716 stars today) — plugin to invoke Codex from inside Claude Code for code review and task delegation. claude-code codex GitHub
  • ogulcancelik/herdr (706 stars today) — terminal-native agent multiplexer for running multiple AI agents concurrently. agentic-ai GitHub

Other

  • sama is in early-stage talks to give the US government a 5% equity stake in OpenAI, worth ~$42.6B at its $852B valuation. The proposal, modeled on Alaska’s oil-wealth fund, has reportedly been raised with Trump, Lutnick, and Bessent; it would require an act of Congress to formalize. CNBC · Bloomberg

  • Court emails released this week from Anthropic’s DOD lawsuit detail the Pentagon’s demand that Anthropic accept autonomous weapons deployment as a condition of contract renewal. CEO Dario Amodei refused on reliability grounds; the dispute escalated to a Trump executive order in February banning federal use of Anthropic products. Anthropic statement · Wikipedia