Skip to content
Source · Daily Brief

AI Morning Brief — 25 July 2026

Anthropic’s Claude Opus 5 dominated the overnight cycle, setting a new state-of-the-art on ARC-AGI-3 at 30.2 percent — nearly four times the previous best — while more than two dozen companies published a joint open letter urging Washington to protect open-weight AI models from broad regulatory restriction.

Top stories

  • Claude Opus 5 launches. Anthropic released Opus 5, positioning it as delivering near-Fable 5 intelligence at half the token price ($5 per million input, $25 per million output). The model debuted at number two on the Vals Index (74.8% versus Fable 5’s 75.1%), set a new ARC-AGI-3 record by solving previously unbeaten puzzle environments through algebraic notation, and immediately integrated into AWS Bedrock and GitHub Copilot. Its system card notes classifiers intervene roughly 85% less often than on Fable 5. via The Verge; via GitHub Blog
  • Industry backs open weights. NVIDIA CEO Jensen Huang joined Meta, Microsoft, IBM, Palantir, Hugging Face, YC, a16z, Mistral, Mozilla, and fifteen other organisations in signing “Open Weights and American AI Leadership,” urging policymakers to avoid premature restrictions and to distinguish legitimate model distillation from IP misappropriation. OpenAI also appeared on the signatory list via a Microsoft webpage. via AI News
  • FLUX 3 released. Black Forest Labs shipped FLUX 3, a multimodal flow model at 5B and 14B parameter scales covering image, video, audio, and action-prediction within a single architecture. The lab claims it outperforms Seedance 2.0 and Grok Imagine on generation quality, and released a companion robotics model, FLUX-Mimic, for video-action tasks. via Black Forest Labs
  • Midjourney ships V8.2 and acquires Co-Star. Midjourney released V8.2 as its new default model — described as more creative and better at personalisation — and simultaneously announced the acquisition of astrology app Co-Star, bringing its CEO Banu Guler aboard as Chief Design Officer. via TechCrunch
  • Kimi K3 distillation row escalates. The White House reportedly accused Moonshot AI of copying Anthropic’s Fable model at scale to build Kimi K3. Separately, UK and US AI safety institutes found K3 scored 32 percent on the ExploitBench offensive-cyber benchmark versus 76 percent for leading US models, with the gap attributed partly to distillation constraints on safety-sensitive capabilities. via TechCrunch

Who shipped

Anthropic owned the cycle with Opus 5, which is now the top-ranked model on Artificial Analysis’s intelligence leaderboard. Black Forest Labs shipped FLUX 3, a meaningful step toward unified multimodal generation in a single open model. OpenAI continued its product push: a Health feature now connects Apple Health data to ChatGPT for US users, and ChatGPT voice mode reached the desktop app with Codex workflow integration. Midjourney shipped V8.2 and broadened its product scope with the Co-Star acquisition. xAI updated Grok Build to v0.2.112 and released a Grok add-in for Google Workspace covering Sheets, Docs, and Slides.

Open-source pulse

Hugging Face released The Stack v3, the largest open code dataset to date at 114 TB, with deduplication and PII redaction applied. AMD published Instella-MoE-16B-A3B on Hugging Face. Laguna S 2.1 received an update addressing early reasoning loop issues, with community benchmarks showing competitive 120B-class performance when run at Q8 quantisation.

Money, infra & hardware

Prentis, a new AI lab co-founded by Reid Hoffman and Marc Pincus, is in talks to raise $100 million, betting that automating routine computer tasks will overtake coding as AI’s biggest commercial use case. Cognition acquired conversational-AI startup Poke in a low-nine-figure deal to bring Poke’s personality model to its Devin coding agent. Stripe is reportedly in talks to acquire model-routing marketplace OpenRouter at a $10 billion valuation. The White House announced $5 billion in Genesis Mission AI-driven science grants. AMD‘s new MI455X GPU carries 432 GB of HBM4 — more memory than five H100s combined.

Quiet corners

Google/DeepMind posted no model releases overnight, and commentary suggests Gemini 3.5 Pro faces a difficult competitive position. DeepSeek was absent from primary announcements, though distillation accusations kept it in the conversation.

By the numbers

  • 339 stories in 24 h across 40+ sources
  • Most-mentioned model: Claude (Opus 5)
  • Most-mentioned lab: Anthropic
  • Notable absences: Google/DeepMind, DeepSeek

Compiled by AI Feed’s editor from all 339 headlines published on 25 July 2026.