Here’s a strange sentence to sit with: Google is already deep into training its next flagship AI model, and the one before it still isn’t out. On July 21, Google shipped three new members of the Gemini family — but the one everyone actually asked for, Gemini 3.5 Pro, was nowhere on the list. Instead, buried in the announcement, Google confirmed that pretraining for Gemini 4 is already underway, calling it the company’s “most ambitious training run” yet. It’s the AI equivalent of a restaurant telling you your appetizer is still in the kitchen while the chef proudly announces next season’s menu.
What Actually Shipped on July 21
Google didn’t come away empty-handed — it released three new models, all in the smaller, faster Flash family. Gemini 3.6 Flash cuts output token usage by about 17% versus its predecessor, at $1.50 per million input tokens and $7.50 per million output tokens. Gemini 3.5 Flash-Lite is built for raw speed, pushing 350 output tokens per second at a bargain $0.30/$2.50 per million tokens. And Gemini 3.5 Flash Cyber is the most unusual of the three: a cybersecurity-specialized model restricted to governments and vetted partners, because its offensive capabilities are considered too risky for general release. None of these are the flagship, though — they’re the reliable, workhorse tier, not the model meant to compete head-to-head with GPT-5.6 or Claude’s top-end releases.
So Where Is Gemini 3.5 Pro?
Sundar Pichai promised at Google I/O back on May 19 that 3.5 Pro would land “within a month.” That month came and went. Then another one did too. Reporting around the delay points to a specific culprit: the flagship is reportedly months behind schedule because Google keeps working to improve its coding performance before letting it out the door — coding benchmarks have become the battleground metric that AI labs get judged on most publicly, so shipping a flagship that loses that fight isn’t really an option anymore. Google’s official line is that it will ship “as soon as it is ready,” and for now it’s living in exclusive partner testing, visible to select companies but not to the public.
The Interesting Part: Why Announce Gemini 4 At All?
This is the detail that makes the story more than just “big company misses a deadline.” Announcing that you’ve started pretraining a model that’s likely a year or more from release, while your current flagship is stuck in limbo, isn’t normal PR behavior — it’s usually a sign a company feels it needs to say something to keep the narrative from being purely negative. Coverage of the announcement has openly framed it as damage control: Google can’t tell you when 3.5 Pro ships, so it tells you what’s coming after it instead. Whether that reassures anyone or just raises the obvious question — “wait, so how far away is 3.5 Pro if you’re already training the next one?” — depends on how charitably you read corporate messaging.
Why This Race Actually Matters to You
It’s easy to treat “which chatbot ships first” as inside baseball, but the timing pressure here is real and competitive. OpenAI launched GPT-5.6 earlier this month, after its own government-requested delay to vet security concerns — we covered that delay in detail here — and that release adds direct pressure on Google to answer with something comparable. When frontier labs are this deadlocked on release timing, the models you actually get access to (the free or cheap tier you’re using in Gmail, Docs, or a coding assistant) tend to lag behind whatever’s winning the headline race. If you’re using Gemini for anything through Google’s enterprise tools, it’s worth knowing you’re likely on 2.5-era infrastructure for a while longer — we broke down what’s actually live right now in our look at Gemini Enterprise.
Quick Tangent: What “Pretraining” Actually Means Here
When Google says Gemini 4 is “in pretraining,” it’s describing the earliest and most expensive phase of building a large language model — the part where the raw model chews through enormous quantities of text, code, and other data to learn general patterns, long before anyone fine-tunes it into a usable chatbot. Pretraining runs for frontier-scale models can take months and burn through tens of thousands of specialized chips running around the clock, which is exactly why the phrase “most ambitious training run yet” is doing a lot of work in that sentence — it’s shorthand for more compute, more data, and a much bigger bill than whatever trained 3.5 Pro. Confirming pretraining has started tells you almost nothing about the release date, though: fine-tuning, safety testing, and the kind of coding-benchmark polishing that’s reportedly holding up 3.5 Pro all happen afterward, and that’s usually where the real delays creep in.
Should You Actually Care Which Model Ships First?
If you’re not building on Google’s API, probably not directly — but the ripple effects show up in places you do notice. Delays at the flagship level tend to slow down features trickling into consumer products like Search’s AI Overviews, Gmail’s smart replies, or Android’s on-device assistant, since those often inherit capabilities from whatever the current top-tier model can do. It’s also a reasonable signal for anyone choosing between AI subscriptions right now: a company visibly stuck between two model generations is, at least temporarily, not the one shipping the newest capabilities to its paid tier. That doesn’t make Gemini worse — 3.6 Flash and the new Flash-Lite are both genuinely competitive on price and speed — but it does mean the flashiest updates are, for now, happening somewhere else.
The Bigger Pattern
Google isn’t alone in racing to announce the next thing before finishing the current one — it’s become something of a habit across the AI industry, where “we’re already training X+1” doubles as both genuine progress update and a hedge against bad press. Anthropic has leaned on similar infrastructure announcements during its own release gaps, and we looked at how that plays out around chip supply and IPO chatter in our Anthropic-Samsung coverage. The pattern says less about any one company’s engineering and more about how thin the margin for public patience has gotten in this industry — a missed month used to be a rounding error; now it’s a headline.
Deixe um comentário