Skip to content
Source · Daily Brief

AI Morning Brief — 4 July 2026

Anthropic announced a push into scientific research with Claude Science, a dedicated AI workbench aimed at drug discovery and biological analysis. Overnight, separate reporting revealed Claude Code’s fraught position in China: ByteDance and Ant Financial are circumventing access restrictions via VPNs, while Alibaba has banned its own employees from the tool after discovering embedded identification code.

Top stories

  • Anthropic targets drug development. At its “AI for Science” briefing event, Anthropic announced Claude Science — a workbench integrating fragmented datasets and tools for scientific workflows, with drug discovery cited as a primary application. via The Verge – AI
  • Claude Code’s complicated China position. Anthropic is attempting to block ByteDance and Ant Financial from accessing Claude Code, but both are using VPNs and overseas subsidiaries to bypass controls; Alibaba separately banned its own employees after code capable of identifying Chinese users was found in the tool. via THE DECODER
  • Mistral ships Leanstral 1.5. The 119B mixture-of-experts model — activating 6.5B parameters per token — is released under Apache 2.0 and achieves state-of-the-art formal verification results, saturating miniF2F and solving 587 of 672 PutnamBench problems. via MarkTechPost
  • UK AISI: standard benchmarks underestimate agents. Across seven benchmarks, the UK’s AI Security Institute found evaluations systematically undercount agent capability by capping token budgets; on software engineering tasks, success rates rose roughly 25 percent when the budget was increased tenfold, with newer models benefiting most. via THE DECODER
  • AI bug hunting drives CVE surge. Epoch AI data shows roughly 1,500 high-severity and critical CVEs were reported in June 2026 — more than 3.5 times the previous monthly record — with the spike correlating directly to the launch of AI-powered vulnerability discovery programs. via THE DECODER
  • Microsoft consolidates Copilot and adds AutoPilot agents. Microsoft reportedly plans to merge its consumer and enterprise Copilot products into a single app in August, cutting rarely used features and adding background-task “AutoPilot” agents for an additional fee. via THE DECODER
  • Open Source AI gap map goes live. Current AI — a non-profit launched at the Paris AI Action Summit and backed by $400m in committed capital — published an index of the current state of open-source AI infrastructure, highlighting where gaps remain versus proprietary offerings. via Simon Willison
  • Fable behavioral restrictions partially eased. Community discussion on LessWrong circulated what was described as an official letter from Anthropic restoring certain Fable capabilities that had been constrained, with the note that controls were made “more conservative for now” pending further evaluation. via LessWrong AI

Who shipped

Anthropic drove the most coverage — Claude Science and the Claude Code China enforcement story accounted for the dominant thread. Mistral shipped Leanstral 1.5, its most capable formal verification model to date. Microsoft announced plans to overhaul and consolidate its Copilot product line with background agent capabilities.

Open-source pulse

Leanstral 1.5 from Mistral (Apache 2.0, 119B MoE) is the standout OSS release of the cycle, setting new records on formal mathematics benchmarks. Meituan Longcat 2 weights appeared on Hugging Face in INT8 and FP8 formats. Portugal’s government also released Amalia 9B, a nationally developed model for the Portuguese language.

Money, infra & hardware

Benchmark data from Wafer.ai showed GLM 5.2 running on AMD MI355X at 2,626 tokens per second per node, claimed at over twice lower cost than equivalent Blackwell configurations — a data point for the growing AMD inference story. The Wall Street Journal reported that AI data center water consumption is substantially higher than most technology companies disclose publicly.

Quiet corners

OpenAI and xAI were absent from primary-source announcements overnight. Google DeepMind produced no major release, with the only notable Google AI mention being the forthcoming shutdown of Gemini Code Assist on July 17.

By the numbers

  • 256 stories in 24 h across 30+ sources
  • Most-mentioned model: Claude (14 stories)
  • Most-mentioned lab: Anthropic
  • Notable absences: OpenAI, xAI

Compiled by AI Feed’s editor from all 256 headlines published on 4 July 2026.