Morning brief

Sunday, July 5

4 people · 0 companies · 6 topics

TL;DR Claude Sonnet 5 launched June 30 as the default agentic model and Fable 5 was fully restored July 1 after a 19-day export-control suspension; Meituan’s LongCat-2.0, a 1.6T open-source model trained entirely on Chinese chips, quietly topped OpenRouter before its June 30 public reveal. (Note: covers June 30-July 5 — gap since last run on June 29.)

Models & research

  • Claude Sonnet 5 launched June 30: 63.2% SWE-bench Pro, 80.4% TerminalBench 2.1 (beats Opus 4.8’s 74.6%), 1M context, $2/$10 MTok intro pricing through Aug 31; default for Free/Pro claude.ai users as of July 1. TechCrunch

  • Fable 5 and Mythos 5 fully restored July 1 globally after 19-day suspension triggered by an Amazon researcher jailbreak that coaxed the model into generating exploit code. Anthropic added a classifier blocking that prompt vector >99% of attempts; Pro/Max/Team get Fable 5 up to 50% of weekly limit through July 7, then via usage credits. Anthropic · 9to5Mac · Hacker News

  • Meituan open-sourced LongCat-2.0 on June 30: 1.6T MoE, MIT license, first model at this scale trained entirely on 50,000 domestic Chinese ASICs with no NVIDIA. Scored 59.5% SWE-bench Pro (vs. GPT-5.5’s 58.6%). Had been quietly topping OpenRouter under alias “Owl Alpha.” Full weights not yet posted on Hugging Face. VentureBeat · HuggingFace

  • elonmusk confirmed June 28 that Grok 4.5 (1.5T params, V9 architecture, supplemental Cursor training data) is in private beta at SpaceX and Tesla. Musk claims “close to or exceeding Opus” but no independent benchmark exists and there is no public API. xAI says it targets a new model monthly for the rest of 2026. Cryptobriefing

  • Meta AI chief Alexandr Wang told staff internally that Watermelon, the next model after Avocado and using 10x more compute, now matches GPT-5.5 on current evals. No benchmarks named, no third-party confirmation. Benzinga

  • GPT-5.6 Sol coming to Cerebras at up to 750 tok/sec in July; currently limited to select customers as capacity expands. OpenAI

People

  • simonw posted July 3: best Fable tip so far is telling it “For all coding tasks use your judgement to decide an appropriate lower power model and run that in a subagent” — noticeably slows drain on the weekly Fable limit. x.com/simonw

Tools & dev

  • claude-code defaults to Sonnet 5 as of v2.1.197, adds org-level default model setting in admin console, and ships two new admin analytics tabs covering developer usage and cost value. Releasebot

  • OpenAI and Broadcom unveiled Jalapeño on June 24: OpenAI’s first custom ASIC, reticle-sized inference chip co-designed in nine months using OpenAI’s own models. Engineering samples running GPT-5.3-Codex-Spark at target perf/watt; deployment targeted for late 2026. gdb told CNBC chips were designed end-to-end using AI assistance. TechCrunch · CNBC

OSS

  • GitHub trending today: openai/codex-plugin-cc (+1,519 stars) — Codex integration plugin for Claude Code enabling cross-tool code review. GitHub · Zackriya-Solutions/meetily (+1,409) — self-hosted AI meeting assistant using Parakeet/Whisper and Ollama, no cloud required. GitHub · usestrix/strix (+1,121 Python stars today) — open-source AI pentesting tool. GitHub

Other

  • Pentagon court documents unsealed July 2 show Anthropic was blocked from DoD prime and subcontracts after refusing a clause permitting “any lawful use” of Claude, which Anthropic said implicitly authorized fully autonomous weapons targeting. Contract cancellations run on a 180-day wind-down. TechTimes · Wikipedia

  • OpenAI proposed giving the US government a 5% equity stake (~$42.6B at $852B valuation) structured as an Alaska Permanent Fund-style sovereign wealth vehicle that would ask other top AI labs to match. Described as “conceptual” and would require an act of Congress; no lab has agreed. Bloomberg · Forbes

  • Anthropic actively cracking down on Chinese companies (Ant Financial, ByteDance) bypassing geographic blocks via Singapore subsidiaries and VPN expense reimbursement; now integrating Persona for government ID checks on high-volume API accounts suspected of proxy routing. TechStory