DeepSeek closed out the week by dropping V4 in two preview models — Pro and Flash — reclaiming the open-weights crown from Kimi and GLM and undercutting every Western frontier on price. OpenAI shipped GPT-5.5 as the next step toward its "super app," Google Cloud Next unveiled two new TPU generations and the Gemini Enterprise Agent Platform, and Anthropic locked in multiple gigawatts of TPU compute through a Broadcom-and-Google deal. Microsoft's Agent Framework 1.0 quietly became the most credible production-grade orchestrator outside the big labs, while CBS News broke that unknown parties had accessed Claude Mythos via a third-party vendor — the first visible crack in Project Glasswing.
Practitioners spent the week wringing Opus 4.7 for everything it's got. Product Compass published the most thorough walkthrough yet of Adaptive Thinking and the new effort levels, Anthropic followed up with official best-practices guidance for pairing Opus 4.7 with Claude Code, and DataCamp's benchmark harness tested how memory and effort trade off in real tasks. MCP content is maturing too: a wordpress-hosted tutorial covers progressive discovery and the three-layer agent stack, Honeycomb's second MCP-plus-Canvas post shows the pattern in production, and a Medium piece distills six hard-won Opus 4.7 tips from inside the Claude Code team.
On the thinking side, MIT Tech Review's state-of-AI chart pack is the reference for macro context, Crunchbase confirmed a record $300B flowed into startups in Q1 — almost entirely AI — and IEEE Spectrum walked through Stanford's AI Index 2026. Understanding AI explains why Meta's Muse Spark means Meta is back in the LLM race, and Foreign Policy explains why governments are reading Project Glasswing as a policy event. For discovery: Simon Willison's DeepSeek V4 teardown is the clearest practitioner read of the release, The Zvi's second Mythos deep-dive is the governance companion, Neural Maze sketches a solo dev's 2026 roadmap, and Bismarck Analysis argues Mistral will benefit most from the coming compute surplus. If you only read one thing this week, make it Simon Willison on DeepSeek V4 — the pricing table alone reframes every model-selection conversation you'll have next week.