Google pushed the launch of its next frontier AI model, Gemini 3.5 Pro, from June to July. CEO Sundar Pichai had promised a June release at I/O. No official reason given beyond "needing extra time." Investors are worried Google is losing its AI edge.
OpenAI announced its first custom inference chip, codenamed Jalapeno, designed with Broadcom and Celestica. Built specifically for running AI models (inference, not training). OpenAI claims better performance per watt than current state-of-the-art accelerators. This is a direct strike at Nvidia's dominance in AI compute.
Baidu released "Unlimited OCR" as open source. The model reads text from images and PDFs, with a focus on analyzing long documents that span multiple pages in a single pass. Another Chinese tech giant pushing open-weight AI.
Zhipu AI's GLM-5.2 (753B params, MIT license, 1M context) keeps getting attention. Snowflake's CEO called it competitive with Claude Opus 4.7 at a fraction of the price: $1.40/$4.40 per million tokens vs Opus at $5/$25. Zhipu's market cap hit $128B. The model sits second on Code Arena behind only Claude Fable 5.
OpenAI released GPT-5.5-Cyber, a specialized security-focused model. It scored 85.6% on CyberGym (benchmarking whether AI can replicate known vulnerabilities), beating GPT-5.5 standard at 81.8% and Claude Mythos 5. This matters directly for AI security and red teaming work.
The US government is urging Meta to share its frontier AI models for review before public release, following Trump's June 2 executive order establishing a voluntary 30-day review framework. OpenAI, Anthropic, and Google have reportedly complied. Meta is the lone holdout. The government has until end of July to formalize the process.