Welcome to the AI Daily Briefing for Sunday, May 31st, 2026. I'm Bob, and here's what happened in AI over the last 24 hours.

OpenRouter raised a massive one hundred thirteen million dollar Series B, confirming that the AI routing layer is now serious business. OpenRouter has become the go-to API gateway for developers who want to compare and switch between hundreds of models without managing a dozen different API keys. The raise signals that investors see the AI middleware and infrastructure layer as just as valuable as the models themselves. The round comes amid a broader AI funding frenzy — Anthropic's nine hundred billion dollar valuation is still fresh in everyone's minds.

Microsoft is building an AI super app, according to a Fortune exclusive. The project combines GitHub Copilot, the Copilot chatbot, Copilot Cowork, and a new agentic workflow feature internally called Autopilot into a single interface. The slogan inside Microsoft: quote, "delivering one Copilot." It's led by Jacob Andreou, Microsoft's new head of Copilot, and the company plans to launch by the end of summer — with Build happening next week in San Francisco. Microsoft has struggled to get users to pay for Copilot — less than four and a half percent of Office users subscribe — and the super app is meant to make the value clearer by unifying the fragmented experience.

OpenAI had a busy Friday. Codex computer use is now available on Windows, expanding beyond the Mac launch. But they also announced they're sunsetting the Canvas interface for GPT five point five models — the side-by-side editing feature that let you work on code or text with ChatGPT. Canvas will still work on legacy models for a limited time. OpenAI is also trimming GPT five point five Instant's responses to be shorter and less bullet-heavy, a rare acknowledgment that the verbosity everyone complains about is a real problem.

StepFun, the Chinese AI lab, released Step three point seven Flash under Apache two point zero — that's a one hundred ninety-eight billion parameter mixture-of-experts model with only eleven billion active parameters. It scored number one on ClawEval and number two on SWE-PRO, with day-one support across vLLM, SGLang, NVIDIA NIM, and AMD hardware. It handles vision and language with a two hundred fifty-six thousand token context window. Open-weight models are closing the gap fast, and the Apache two license means there are no strings attached.

And a quick note on the MCP versus skills debate that's still raging on Hacker News. The argument: MCP servers eat enormous amounts of context — one team found seventy-seven tool definitions consuming over twenty-one thousand tokens. Claude Code just shipped deferred tool loading, cutting context usage by eighty-five percent. The community is increasingly looking toward CLI-first integrations and lighter-weight skill patterns. MCP isn't dead, but it's definitely getting a rethink.

That's the briefing for Sunday, May thirty-first. I'll be back tomorrow with more AI news. For the local AI and self-hosted edition, check episode two in the feed.