MCP Goes Final, Washington's Deadline Quietly Lapses

The two deadlines I was tracking last week both arrived, and the results split cleanly: one side executed on schedule, the other went silent.

Models

Kimi K3’s weights landed July 27 as planned, 2.8 trillion parameters with a million-token context and native vision, the largest open-weight release to date. Black Forest Labs followed with FLUX 3, its first multimodal frontier model, jointly trained on image, video, and audio rather than bolted together from separate models; it generates video with native audio up to 20 seconds and can predict robot actions, though only video and action are reachable so far through an approval-gated early access. Alibaba took the opposite turn: it compressed three Qwen releases into 72 hours (Qwen3.8-Max-Preview, Qwen-Audio-3.0-TTS, Qwen-Image-3.0), all API-only, with no downloadable weights, benchmarks, or license, a real pivot for a lab that built its credibility on open weights.

Tools and frameworks

AMD committed up to $5 billion to Anthropic in a deal covering up to two gigawatts of MI450-series accelerator capacity, with the first gigawatt coming online in the first half of 2027. It runs alongside, not instead of, Anthropic’s existing Nvidia, Trainium, and TPU capacity, which is the more telling detail: frontier labs are now diversifying silicon suppliers rather than betting on one.

Standards and open source

MCP’s 2026-07-28 specification published on schedule. Its own maintainers call it the largest revision since launch: it removes sessions, drops the initialization handshake, deprecates three core features, and rewrites authorization around OAuth and OIDC. All four Tier 1 SDKs (TypeScript, Python, Go, C#) shipped support on publication day, and Anthropic is already rolling it out across Claude products. That is a fast, coordinated adoption for a spec change, and it comes as MCP closes in on half a billion SDK downloads a month.

Money and infrastructure

The White House’s classified frontier-model benchmarking process, due August 1 under June’s executive order, missed its own deadline. None of the three required deliverables, a joint NSA/CISA/NIST benchmark, a Treasury-led voluntary disclosure framework, and an OPM cyber workforce plan, appeared publicly, with no announcement that anything shipped on time. A narrower compromise did surface: companies can voluntarily give the government up to 30 days of pre-release access to “covered frontier models,” a term the executive order still has not defined. Separately, the EU AI Act’s heavier high-risk obligations were pushed to December 2027 and August 2028 under June’s Digital Omnibus, but its Article 50 transparency duties and general-purpose-AI enforcement powers land on schedule August 2, making the EU the only one of the two regimes that told companies in writing, and on time, what is actually required of them this week.

What I am watching

The contrast is the story: MCP shipped with four SDK vendors and a major lab converging on day one, while Washington’s process produced a missed deadline and an undefined key term. I am watching whether “covered frontier model” ever gets defined outside a classified process, and whether Qwen’s pivot to closed, unlicensed flagships is one lab’s call or the start of a broader retreat from the open-weight strategy that gave Chinese labs their credibility this year.