// NODE-02 · NEURAL FEED · DAILY TRANSMISSION //
June 26, 2026 · WINDYVIEWS.COM
Google launched Gemini 2.5 Pro with a new Deep Think reasoning mode on June 22, posting 82.4% on GPQA Diamond — graduate-level physics, chemistry, and biology questions. That edges out Anthropic's Fable 5 (79.1%) and OpenAI's GPT-5.5 (76.3%), putting Google back at the front of the frontier reasoning race.
On June 9, Anthropic released Claude Fable 5 and Mythos 5, a capability tier above Opus 4.8. Days later, a June 12 emergency directive from the U.S. Department of Commerce barred foreign-national access on national-security grounds, and both models remained offline as of June 22 — a rare case of a flagship model launch colliding head-on with export policy.
Noam Shazeer — a co-author of the 2017 "Attention Is All You Need" paper that introduced the Transformer — is leaving Google DeepMind to join OpenAI as Lead for Architecture Research, owning the core neural-network design across OpenAI's models. It's one of the most consequential researcher moves of the year.
Nobel laureate John Jumper, creator of DeepMind's AlphaFold, is leaving Google DeepMind for Anthropic. Paired with the Shazeer move, it underscores how aggressively OpenAI and Anthropic are now poaching the field's most decorated researchers away from Google.
SpaceX closed the largest startup acquisition in history, buying AI coding company Cursor for $60 billion. The deal signals how strategically valuable AI-native developer tools have become — and pulls a major coding-assistant player into Elon Musk's orbit.
Sensor Tower's 2026 State of AI Report puts ChatGPT's share at 46.4% by late May — its first dip under 50%. Gemini climbed to 27.7% and Claude reached 10.3%, evidence that the consumer assistant market is genuinely fragmenting rather than consolidating around a single winner.
On June 13, China's Zhipu AI open-sourced GLM-5.2, a large Mixture-of-Experts model with a usable 1-million-token context window, shipped under the permissive MIT license. It's a strong signal that open-weight frontier models with very long context are staying competitive with closed labs.
OpenAI launched GPT-5.5-Cyber, a model tuned for cybersecurity tasks, scoring 85.6% on the CyberGym benchmark and edging out Anthropic's Mythos 5. The release reflects a growing push toward domain-specialized frontier models rather than one general system for everything.
Morgan Stanley projects global AI-linked debt will nearly double to $570 billion this year, driven by the scale of infrastructure financing across hyperscalers and frontier labs. Meanwhile U.S. federal AI obligations hit $7.2 billion (up 966% from 2024), as the buildout's capital intensity becomes a story in its own right.
The frontier is a three-way race again: Google's Gemini 2.5 Pro Deep Think retook the reasoning lead just as Anthropic's Fable 5 / Mythos 5 launch got frozen by U.S. export controls — policy is now a first-order product risk. Talent is consolidating toward OpenAI and Anthropic (Shazeer and Jumper both leaving DeepMind), open weights stay competitive via Zhipu's MIT-licensed GLM-5.2, and the money side is getting heavy — a $60B Cursor acquisition and Morgan Stanley's $570B AI-debt forecast show the capital, not just the models, is now the headline.