Skip to content
Source · Daily Brief

AI Morning Brief — 5 July 2026

Anthropic drove the overnight cycle with two moves: the June 30 beta launch of Claude Science, a multi-agent scientific workbench for genomics and proteomics pipelines, and an announcement that it is launching its own drug discovery program targeting neglected diseases. A disclosed session leakage issue in Claude Code and Alibaba’s reported ban on the tool added an enterprise caution note to an otherwise active Anthropic week.

Top stories

  • Anthropic releases Claude Science in beta. A coordinating agent delegates work to domain specialists across genomics, proteomics, and cheminformatics; a reviewer agent flags citation errors; and every output figure ships with its exact code, environment, and full message history. via MarkTechPost
  • Anthropic enters drug discovery. The company announced it will run its own development program for neglected diseases the pharmaceutical industry deems unprofitable. Coverage cited Novartis CEO Vas Narasimhan’s estimate that AI could compress the development timeline from twelve years to seven or eight and double the clinical success rate. via THE DECODER
  • Alibaba reportedly bans Claude Code. Alibaba has classified Claude Code as high-risk software and is prohibiting employees from using it, according to reporting published July 4. No public explanation was given. via TechCrunch – AI
  • Mistral releases Leanstral 1.5 for formal verification. The open-source model targets Lean 4 and, beyond benchmark performance, scanned 57 open-source repositories and found five previously unknown bugs. via THE DECODER
  • Claude Opus 4.8 invents extra tool arguments. Developer Armin Ronacher documented cases where Opus 4.8 appended fabricated fields to nested tool calls — structurally invalid but often substantively correct — consistent with excessive on-policy RL, per follow-up researcher commentary. via Hacker News (front page), via Simon Willison
  • Claude Code session leakage disclosed. A GitHub issue on the anthropics/claude-code repository, reaching 31 points on Hacker News, reported potential session and cache leakage between workspace instances or consumer accounts. via Hacker News (front page)

Who shipped

Anthropic led the window with Claude Science Beta, its drug discovery initiative, and developer prompting guidance for Fable 5. Mistral shipped open-source Leanstral 1.5. On the infrastructure side, llama.cpp pushed two patch releases, b9871 and b9873, fixing CPU concat operations for quantized types and a KV rotation buffer guard.

Open-source pulse

Mistral’s Leanstral 1.5 is the most concrete OSS release of the window, with demonstrable real-world bug-finding across 57 public repositories. Meituan’s LongCat 2.0 weights landed on Hugging Face, described as the largest publicly known pretraining run on non-Western silicon, drawing attention from observers tracking the Huawei accelerator stack.

Quiet corners

OpenAI was present only through a Greg Brockman vision piece about a future with no software interfaces. xAI and Meta posted no primary product announcements across the window.

By the numbers

  • 161 stories in 24 h across 15 sources
  • Most-mentioned model: Claude
  • Most-mentioned lab: Anthropic
  • Notable absences: xAI, Meta, Cohere

Compiled by AI Feed’s editor from all 161 headlines published on 5 July 2026.