Gemini 3.5 Pro Is Three Weeks Late, and 'Announced' Has Stopped Meaning Anything

Everyone does this now

Pick your lab, and you can find the same move this quarter alone. Anthropic shipped Opus 4.8 in late May while signaling that its Mythos-class models would arrive "in the coming weeks," a phrase that does a lot of load-bearing work. OpenAI announced its own confidential S-1 filing, which is to say it announced a thing whose defining feature is being confidential, because "we expect it to leak so we're just announcing it." Microsoft unveiled seven models at Build, several of which exist mostly as waitlists outside the Copilot picker.

The keynote and the product have come apart. An announcement used to mean "you can have this now," then it meant "next month," and now it means "we have benchmarked this internally and would like the news cycle today, please."

Why the labs can't help themselves

I'm sympathetic to the mechanics, having watched them up close in policy fights. Three forces make announcement inflation nearly inevitable.

Benchmarks leak, so a lab that waits for general availability risks its rival framing the story first. Talent and capital read momentum, and in a quarter when two labs filed S-1s and a third went public inside a rocket company, looking slow is expensive in ways a CFO can measure. And demos are cheap relative to deployment: serving two million tokens of context to everyone is an infrastructure project, while showing it to a keynote audience is a Tuesday.

All true. All reasonable. And the cumulative effect is a sector where the public can no longer tell the difference between products and press releases, in the same season the industry asks public markets and Congress to take its safety commitments on faith.

The part that actually matters

Here's why I'd rather pick this fight than wave it off as marketing weather. The Great American AI Act draft would have frontier labs publish safety frameworks and sit for semi-annual audits. The entire premise of that regime is that what a lab says in a document corresponds to what is actually running.

An industry that habitually announces things that don't ship is training regulators, courts, and the public to discount its statements. You cannot spend Monday teaching everyone that your words are aspirational and Thursday asking them to treat your safety framework as binding. Credibility is one account; every keynote writes a check against it.

A model exists when I can pay for a token

Gemini 3.5 Pro will ship, probably within weeks of the prediction markets' guess, and it will probably be excellent. Google's sin here is small and common, which is precisely the point. The fix is also small: date-stamped availability commitments in the announcement itself, and the discipline to skip the keynote when the answer is "we don't know yet."

Until then, my house style: a model exists when I can send it a token and pay for the reply. Everything else is a trailer.