Google no-showed its own launch party

Good morning…

it's Monday, you're reading Main Street AI and I'm your host, Jack. Quieter week on paper. Much louder week if you read the footnotes.

Our featured story: July 17 came. July 17 went. Gemini 3.5 Pro did not show up to its own party, and the new whisper is August. Meanwhile, Microsoft's CEO spent Wednesday telling his own engineers that the model he put $5 billion behind is "editorially controlled," China quietly shipped the biggest open model in history, Apple's heist-movie lawsuit picked up a clerical-error subplot, and the Fable meter — delayed twice by angry replies — finally, actually, probably clicked on this morning.

Last Monday I asked who has the worse July: OpenAI's lawyers or Google's. Google answered for you.

But before we start…

Seven days ago I gave you three live questions. Does Google ship Gemini 3.5 Pro on the 17th? No. Does Fable survive on subscriptions past the 19th? No. Which lab drops the next frontier model? Moonshot, out of Beijing, with 2.8 trillion open parameters.

Three questions, three resolutions, one week. This is why prediction markets and this newsletter are the same hobby. The live one now: when does OpenAI officially announce its IPO? Per CNBC, Kalshi traders put it at 59% by March 2027 and only about one-in-three before New Year's, after reports Sam is holding out for a trillion-dollar sticker. If you have a view on Sam's patience, that is now a tradeable opinion.

(Usual reminder: I'm a guy with a newsletter, not your advisor. Only trade what you understand.)

Trade on real-world events you already follow, from elections and inflation to sports, tech, and more. Choose “Yes” or “No” based on what you think will happen and see how your prediction plays out.

Pick your market and start trading what’s next.

Trade $25, Get Up to $500

_{Bonus credit varies from $15 to $500. Terms apply.}

But anyways, let's get into it:

The boring details are…

July 17 was the date. The whole internet circled it, I printed it here last Monday, and the day arrived with… nothing. No Gemini 3.5 Pro. As of this weekend, Pro is still parked in a limited Vertex AI preview, and the new whisper number is August.

Recap for new readers: Google announced the 3.5 family at I/O on May 19, promised Pro for June, reportedly scrapped the entire base model after finding structural problems in agentic tool-calling, restarted pretraining from scratch, slipped to July, and has now slipped past its own rumored date too. The leaked specs remain gorgeous — a 2 million token context window, a built-in "Deep Think" reasoning mode. Leaked being the operative word, because none of it exists in public.

Three things.

First, the open door. GPT-5.6 had to sit through extra Commerce Department testing to go fully public. Fable 5 spent 19 days suspended under an export-control directive and came back wearing a meter and a safety classifier. Gemini is the one frontier model with no government anchor tied to its ankle — and it's the one model that doesn't exist. You cannot write comedy this clean.

Second, the stakes went UP this week. Reporting now says a custom Gemini model will power Apple's rebuilt Siri. Remember the landlord thesis from June: Apple wins by being the device every model runs on. The landlord just signed a lease with the tenant who keeps missing the move-in date.

Third, credit where due: Gemini 3.5 Flash has quietly carried Google since May, beating the older 3.1 Pro on coding and agent benchmarks at $1.50 in / $9 out. The cheap little brother is doing the flagship's job. That's either a lovely story about efficiency or a damning one about the flagship. Ask me in August. Apparently.

You had one job. Part four.

Wednesday, internal Copilot all-hands. Microsoft CEO Satya Nadella, per a transcript CNBC got its hands on, went after Anthropic's Fable 5. His phrase: "editorially controlled." He said the refusals don't make sense, asked when a creation tool was ever policed like this, and then swung at the whole industry, arguing it can't be that two companies own all the token capital while everybody else rents it.

Now the context, because the context IS the story:

Microsoft has $5 billion in Anthropic. Anthropic has committed a reported $30 billion to Azure. Microsoft already restricted employees' internal use of Fable this year over Anthropic's data-retention policy, and — June 1, you were here — cancelled Claude licences over price. Microsoft stock is down 17% on the year while the Nasdaq is up 11. And Anthropic's annualized revenue reportedly just blew past $30 billion.

So read the play. When a CEO trashes a partner he's invested in, to his own engineers, in remarks he knows will leak, that is not a product review. That's a renegotiation performed for an audience. The tell to watch, and industry watchers said the quiet part out loud: does Anthropic quietly loosen the refusals, or does Microsoft lean harder on its in-house MAI models next time Copilot gets blocked? Either way, Satya won the meeting.

Zoom out. In three weeks the safety brand has been publicly torched by Alibaba (uninstall everything), by its own user base (the meter riots), and now by its second-biggest patron. Every complaint doubles as a negotiation. Nobody in this story is a bystander. Still.

Two delays. Two Mondays of me writing about the delays. And now: as of this morning, July 20, the included window is over and Fable 5 runs on usage credits. $10 per million tokens in, $50 out — double Opus 4.8, the most expensive sticker Anthropic has ever put on a generally available model.

The scoreboard, for posterity: original cutoff July 7, panic-extended hours before to July 12, panic-extended again through July 19. No third blink as I hit send Sunday night. If they folded overnight, congratulations — you're reading a collector's item.

Two honest notes. One: Anthropic is still publicly promising Fable comes back to normal subscriptions "when capacity allows," and the capacity excuse looks increasingly real. The Colossus expansion reportedly added 220,000-plus GPUs, and on July 10 Anthropic wiped everyone's Claude Code limits clean after a week where the status page did most of the talking. That's a company shoveling coal into a boiler, not one twirling a mustache.

Two: the temporary 50% Claude Code limit boost — the one everyone read as a Codex response — expired the very same week. Giveth, taketh, meter-eth. The bundled-flagship era is over, and every other lab is quietly thrilled Anthropic went first.

Thursday, July 16: Moonshot AI, Beijing, released Kimi K3. The numbers are absurd. 2.8 trillion parameters — the first open model in the 3T class, ever. A million tokens of context. Native vision. Full weights promised by July 27. Priced at $3 in / $15 out.

Even Moonshot's own launch blog admits K3 still trails Fable 5 and GPT-5.6 Sol. But it beat everything else they tested, and independent scoring puts it far above the open-model pack — with one very funny catch: it is spectacularly verbose. Artificial Analysis burned roughly $2,700 in tokens just benchmarking it. The model thinks like it's paid by the word.

Why you should care: remember Alibaba's Claude Code ban two weeks ago, and my line that the ban was a national advertisement for home-grown tools? The advertisement now has a flagship product. Near-frontier, open weights, undercutting everyone on price, from a company that raised $500 million in January specifically to build it. The "every accusation is a policy argument" season just got its answer from the other side of the Pacific: fine — we'll grow our own.

Update on last week's featured story. Apple's lawsuit claimed OpenAI "never responded" to its February letter about the alleged trade-secret buffet. Per NBC News: emails show OpenAI DID respond in February — and the talks then collapsed after Apple's outside lawyer mixed up the names and email addresses of two OpenAI employees, one surnamed Wang, one surnamed Chang.

The suit itself is unchanged and still radioactive. TechCrunch went through the 41-page filing and surfaced, among other gems, an alleged message from one defendant to a friend still inside Apple: "LOL, I found out I can access the [network storage], so funny." And Bloomberg's read stands — discovery alone threatens OpenAI's device timeline no matter who wins. But "we sent a letter into the void" was Apple's cleanest talking point, and it now carries an asterisk shaped like a paralegal. Both of these companies are run by humans. Exhibit lists cut both ways.

The shovel took a breather. SK Hynix, fresh off the biggest foreign US debut ever, gave the pop back: closed Friday at $154 versus a $168 first-day close, dipping under $146 midweek. Still above the $149 IPO price. Shovels win over years, not weeks. Patience.

SPCX fell again after a Starship test flight was aborted, per CNBC. Second straight rough week for the rocket company that owns a lab. The market keeps grading the rockets, not the models. Cold.

No jokes on this one. xAI sued a user in federal court, alleging he abused Grok to create illegal sexual images of minors — one of the first cases of an AI company suing its own customer over generated content. The precedent, platform as plaintiff, will echo through every lab's terms of service. Watch this.

The backlash found its capital. CNBC spent the week on the fight around Musk's Memphis compute buildout, calling it the epicenter of the anti-data-center movement. The protest thread from last Monday keeps compounding. Tiny, still. Growing, still.

Well folks, that's the week Google missed its own party, Satya negotiated via all-hands, the meter finally clicked on, and Beijing open-sourced a 2.8-trillion-parameter rebuttal to the entire argument.

Keep the replies coming. I read every single one and the best ones end up down here with a shoutout. This week's question: who blinks first — Google actually shipping Pro, or Anthropic taking Fable off the meter? Hit reply, convince me.

See you Monday. Stay sharp out there.

Jack