Seventeen percent fewer tokens is the product
Three efficiency models shipped. The flagship did not. That is not a gap in the roadmap, it is the roadmap.
Google DeepMind released 3.6 Flash, 3.5 Flash-Lite and 3.5 Flash Cyber, with the workhorse cutting token usage by up to 17%. There was no 3.5 Pro.
Where the competition actually is
Token reduction is a direct margin improvement for every customer and a direct cost reduction for the provider, compounding across billions of calls in a way that a benchmark point never does. Shipping three efficiency tiers and no flagship is a statement about which of those two things customers are currently choosing on.
It is the same signal the funding market gave when two inference-serving companies raised one and a half billion dollars each within weeks. Capability has compressed enough that most enterprise workloads are genuinely indifferent between two or three models, while cost differences remain large enough to decide viability.
The number needs its context
Seventeen percent fewer tokens for the same answer is a real efficiency gain. Seventeen percent fewer tokens because the model reasons less visibly is a different thing wearing the same number. Release notes are where that distinction lives, and headlines are where it dies.
This is becoming a general problem with efficiency claims. As compute-per-query turns into a caller-set parameter rather than a model property, every quoted figure needs its configuration attached or it is not comparable to anything.
The tier that should worry you
Flash Cyber is fine-tuned for finding and fixing security vulnerabilities and ships only to governments and trusted partners under a limited-access pilot.
The reasoning is unarguable: a model good at finding vulnerabilities is equally good at finding them for the wrong reasons, and the gap between defensive and offensive use here is intent rather than capability. Restricting distribution is the only lever that does not require solving intent.
But it establishes something. Once capabilities ship by relationship rather than by API key, the questions become which capabilities and who decides — and those decisions are currently made internally, announced as product news, and reviewed by nobody. In the same week that a federal framework kept its criteria private, that is two mechanisms for gating frontier capability by trust rather than rule, neither observable from outside.
TechCrunch — Google releases three new Gemini models — but no 3.5 Pro → · Google AI for Developers — Gemini API release notes → · Google Cloud — Model versions and lifecycle, Gemini Enterprise Agent Platform →