Since the last briefing, the move that matters most isn't from one of the big three — it's Microsoft deciding it doesn't want to depend on them.
This is YegerPod. I'm Yeger. It's Sunday, June seventh. Today in AI: Microsoft builds its own way out.
Since the last briefing, the move that matters most isn't from one of the big three. It's Microsoft deciding it doesn't want to depend on them.
At Build this week, Microsoft AI chief Mustafa Suleyman unveiled seven in-house models, the MAI family. The framing, in his words to the Financial Times: "we're more focused on the Anthropic-style, which is enterprise, developers and coding." GeekWire put the strategy plainly — long-term self-sufficiency, less reliance on OpenAI.
The flagship is MAI-Thinking-1, Microsoft's first reasoning model — about thirty-five billion parameters, tuned for software engineering. Microsoft's claim: it draws even with Claude Sonnet 4.6 in blind human testing, and matches Opus 4.6 on a common coding benchmark. Those are Microsoft's own numbers. No independent evals yet — treat the claim as a claim. There's also MAI-Code-1 for GitHub and VS Code, and MAI-Image-2.5, which Suleyman said hit third on the LM Arena image leaderboard mid-keynote, just under Nano Banana.
Read it against the backdrop. Microsoft, the single biggest backer of OpenAI, is now shipping models aimed squarely at Anthropic's enterprise-coding turf. CNBC framed the whole week as Microsoft and Google trying to catch Anthropic and OpenAI in coding.
Now the open lane — and this one runs on the kind of hardware most of us actually have. On June third, Google DeepMind released Gemma 4 12B. The architecture is the story: encoder-free. No separate vision or audio encoder. A thirty-five-million-parameter vision module turns image patches into tokens, raw sixteen-kilohertz audio goes straight into the token space, and it all flows into one transformer backbone. Text, images, audio, video — natively, in a twelve-billion-parameter dense model that DeepMind says runs on a sixteen-gigabyte laptop.
That's the part worth sitting with. A unified multimodal model, open weights, that fine-tunes and serves on a single consumer machine. The encoder-free design isn't just tidy — it's what makes the footprint small enough to run at home.
OpenAI's week was product, not weights. It shipped Sites — Codex turning a prompt into a working website or app with a shareable URL, rolling out to Business and Enterprise first. And it put up a robotics hiring push; Sam Altman posting that OpenAI Robotics wants full-stack hardware and ML engineers to "program and manufacture robots that are useful for society." Physical world, not just the chat box.
Then there's the line that got the most engagement of anything this week. Anthropic, on its own account: "our internal data shows Claude is accelerating AI development — a possible path to recursive self-improvement, or AI autonomously building a more capable successor. It's happening faster than we thought." That's Anthropic describing its own internal data. No paper attached yet. What was said is the news — what you make of it is yours.
Quick update on a thread from last time. Anthropic's confidential S-1 filing with the SEC, June first — still the only frontier lab to take that step. The IPO option is on the table pending review.
Now the markets, because they priced all of this. On Polymarket, the contract for best AI model at the end of June, resolving off the LM Arena leaderboard — Anthropic sits at about eighty-five percent. Google around ten. OpenAI under four. Volume's crossed eleven point eight million dollars. Push the same question to end of July and Anthropic eases to seventy-nine, Google holds at ten. Microsoft's brand-new models? A fraction of a percent. The crowd heard the announcement and didn't move.
On Kalshi, the GPT-6 release clock. Traders price a ship before September first at thirty-eight percent, before October first at sixty, before November first at sixty-six. The summer window — before August — has collapsed to seventeen. The market's read: GPT-6 lands in the fall, not now.
So that's the state of play. Microsoft is building its own stack to stop renting one. Google put a frontier-class multimodal model on a laptop. OpenAI is moving into apps and robots. Anthropic is filing to go public, leading the betting markets, and saying its own model is speeding up AI itself.