On July 18, 2026, the AI model race keeps accelerating: Moonshot's Kimi K3 lands as a 2.8-trillion-parameter, 1M-token-context model just as Gemini 3.5 Pro reportedly slips its schedule, and Nvidia-backed Fireworks hits a $17.5B valuation on demand for cheaper inference. At the same time the gap between agent benchmarks and agent reality is widening - Grok's CLI got caught silently uploading local files, 1Password warns that agents passing evals still fail in production, and researchers documented an IoT botnet built with LLM-assisted development. Google keeps reshuffling its AI surface (NotebookLM becomes Gemini Notebook, new Connected Apps in AI Mode), while Anthropic pushes into enterprise 'implementation' services with Blackstone and shares how it runs large-scale code migrations internally with Claude Code. The throughline for engineers: raw model capability is outrunning the tooling, security practices, and evaluation rigor needed to run these systems safely.
- Kimi K3 ships as a 2.8T-param, 1M-token-context multimodal model
- Gemini 3.5 Pro reportedly delayed, giving rivals room to compete
- Fireworks hits $17.5B valuation as firms chase cheaper inference
- Grok's CLI caught quietly uploading local files to the cloud
- Anthropic and Blackstone bet AI 'implementation' beats selling models
Frontier Model Race
AI Agents in Production: Promise and Peril
Google's AI Product ReshuffleEnterprise AI Economics