Vercel, Google, and Mistral ship major AI infrastructure updates in same week

Vercel's AI Gateway introduced firewall-style routing rules that let platform teams swap or block models at the credential level without changing application code, reducing model migration to a single config update. Google released Nano Banana 2 Lite, capable of generating 1,000 images in four seconds at low cost, alongside Omni Flash, which enables natural-language video editing within the same API pipeline but is limited to 10-second clips with no audio support. A new suite of MIT-licensed agentic coding models, ranging from 9B to 397B parameters, was released with reinforcement learning training optimized for both solution quality and search scaffolding, supporting 256K context windows. The smaller 9B dense model runs on a single 80GB GPU, making capable agentic coding accessible without multi-GPU infrastructure. Mistral also shipped two production-ready releases in the same cycle, with a text-to-speech offering among them, reflecting a broader industry push toward tighter control over model selection, credentials, and tool integrations.