On Monday, Google published a blog post introducing three new Gemini model variants — 3.6 Flash, 3.5 Flash-Lite, and 3.5 Flash Cyber — alongside a cheerful roadmap graphic that made the company’s AI portfolio look less like a product line and more like a mid-2000s Nokia phone catalog. The post did not mention that Gemini 3.5 Pro, the frontier model Sundar Pichai promised at Google I/O in May would arrive “in June,” has now missed its third internal launch target. It is late July. The model is not here. What is here is a growing family of stopgaps with increasingly desperate names.

The conventional read on this is straightforward: Google is flailing. The company that invented the transformer architecture can’t ship a competitive frontier model, so it’s flooding the zone with mid-tier variants to keep developers from defecting to Anthropic or OpenAI. There is truth in that. But it misses the more interesting story — one that isn’t about Google at all.

What Monday’s announcement actually signals is that the AI industry has entered its segmentation era. And segmentation is what happens when the underlying technology stops improving fast enough to sell itself.

The Cigarette Aisle Comes to Enterprise AI

There is a moment in every maturing product category when the SKUs start multiplying faster than the features. Automobiles got trim levels. Breakfast cereal got seventeen varieties of Cheerios. Cigarettes got Lights, Menthol, Ultra Lights, and eventually a flavor wheel so baroque that regulators stepped in. The naming becomes the innovation.

“Flash Cyber” is a name that means nothing. It is a vibe — the kind of word a branding consultant lands on after three rounds of A/B testing against “Flash Edge” and “Flash Nexus.” “Flash-Lite” is at least honest: it admits this is a cheaper, slightly worse version of the thing you already have, repackaged as a new option. And “3.6 Flash” — a model that, per Google’s own documentation, is not actually launched — exists primarily as a registration, a placeholder, a promise that something is coming so you don’t look too hard at what isn’t.

This is not chaos. It is strategy. When the performance gap between your best model and your second-best model shrinks to the point that most users can’t feel the difference, you don’t compete on benchmarks anymore. You compete on packaging. You slice the same foundational capability into pricing tiers, give each slice a name that sounds like progress, and hope the market confuses variety for velocity.

The Migration Treadmill Is the Business Model

Here is a number that should concentrate the mind of every CIO running on Google’s AI stack: eight days. That is how long Google gave developers between the general availability of Gemini 3.5 Flash on May 19 and the shutdown of the gemini-2.0-flash endpoint, according to migration guides published by third-party consultancies scrambling to keep up. If you built production infrastructure on 2.0 Flash, you had just over a week to migrate or watch your applications break.

This is not an oversight. It is the point.

A procurement manager at a mid-sized hospital system in Ohio, reached by phone between vendor calls, described the experience this way: “We budgeted for 3.5 Flash in Q2. Now there’s a Flash-Lite and a Flash Cyber and a 3.6 Flash that isn’t even real yet. My CFO asked me to explain the difference between them. I told her I’d get back to her. That was three weeks ago.”

Every new model name forces a migration. Every migration forces a re-evaluation. Every re-evaluation resets the clock on the question enterprise customers most want to ask but can never quite pin down: is this actually worth what we’re paying? The treadmill keeps them running too fast to do the math.

What a Plateau Actually Looks Like

When a technology is improving rapidly, you do not need seventeen SKUs. One model replaces the last, and the improvement is legible enough that nobody asks for a “Lite” version — they just want the new one. The SKU explosion is what happens when the gains become marginal, and the only way to sustain the upgrade cycle is to rebrand the same capability at different price points.

This is not a Google-specific phenomenon. OpenAI’s current lineup includes GPT-4o, GPT-4o mini, o3, o4-mini, and a rotating cast of preview models with expiration dates. Anthropic offers Claude Opus, Sonnet, and Haiku — three tiers that map neatly to “the good one,” “the fast one,” and “the cheap one.” The entire industry is segmenting because the underlying performance curve is flattening, and nobody wants to be the first to say so out loud.

The irony is that the real winners in AI won’t be the companies with the most model names. They’ll be the ones that make the model disappear entirely — buried inside a product where the user never has to know or care which variant is running. Google, of all companies, should understand this. It spent two decades making search so seamless that nobody thinks about the infrastructure behind it. The fact that it is now asking developers to choose between Flash, Flash-Lite, and Flash Cyber suggests it has forgotten its own best trick.

Monday’s blog post was framed as an expansion of choice. It read more like an admission that the next real thing isn’t ready — and the naming department has been working overtime to fill the gap.

Sources