Gemini delay, China agent rules and robotaxi expansion lead AI agenda

Gemini 3.5 Pro remains in limited Vertex AI enterprise preview after missing earlier release targets, with Google citing early tester feedback on excessive token consumption, coding gaps and weaker long-horizon reasoning. The delay comes as rivals move ahead with Claude Sonnet 5, Claude Fable 5 and GPT-5.6 previews, raising pressure on Google to show clear gains in long-context retrieval and hard reasoning.

China’s AI companion rules take effect on July 15, prompting ByteDance’s Doubao and Alibaba’s Qwen to shut down humanlike and user-created agent features. Doubao, which has 345 million monthly active users, will keep agent histories in read-only mode until October 15, 2026, while Alibaba has announced no migration path for Qwen users. Tesla has also expanded Robotaxi service to Miami without a safety monitor, making it the company’s fifth US city.

Model benchmarking and infrastructure news underscored a broader shift in the market. Meta released SWE-Together, where Claude Opus 4.8 reached 63% pass@1 on multi-turn coding tasks, while OpenAI’s GeneBench-Pro showed GPT-5.6 Sol Pro scoring 31.5% on expert biology problems. Chinese providers now handle approximately 45% of OpenRouter traffic, with Xiaomi processing 4.21 trillion weekly tokens for a 21.1% share versus OpenAI’s 7.5%.