Google releases Gemini 3.6 Flash, 3.5 Flash-Lite and 3.5 Flash Cyber

Background

Large model providers rarely ship one model at a time. A generation usually appears as a family: a general-purpose model, a smaller and cheaper one for high-volume work, and sometimes a variant tuned for a single kind of task. Buyers rarely choose on quality alone; the practical question is which model is good enough at a price and speed they can afford. Prices are quoted per million tokens — the units of text a model reads (input) and writes (output) — so the two rates are listed separately.

What happened

On 21 July 2026 Google announced three models at once: Gemini 3.6 Flash, Gemini 3.5 Flash-Lite and Gemini 3.5 Flash Cyber^[1].

The main model, Gemini 3.6 Flash, improved on 3.5 Flash in coding and agent benchmarks while using fewer output tokens^[1]. Fewer output tokens matters because output is the slower and usually the more expensive half of a request. Google priced 3.6 Flash at $1.50 per million input tokens and $7.50 per million output tokens^[1]. Google's Gemini release notes for the same day said the model was being brought to all Gemini app users worldwide, so it was not only an API offering but part of the consumer app^[2].

Gemini 3.5 Flash-Lite took the other end of the trade-off. It ran at 350 output tokens per second and cost $0.30 per million input tokens^[1], with output at $2.50 per million tokens^[1]. High generation speed with a low input price points at workloads where a system reads a lot of text and must answer quickly, and where a small error rate is acceptable.

The third model was different in kind. Gemini 3.5 Flash Cyber was fine-tuned to find and fix vulnerabilities, and access was limited to governments and trusted partners in a CodeMender pilot^[1].

Google also said that Gemini 3.5 Pro was still in testing with partners and that pre-training of Gemini 4 had begun^[1].

Key details

The announcement is notable for how the three models split the space. Flash-Lite and Cyber carry the 3.5 version number, while the general-purpose Flash moves to 3.6, so version numbers here follow the tuning of a particular line rather than a single company-wide release.

The price gap is the clearest signal. Input tokens cost five times more on 3.6 Flash than on 3.5 Flash-Lite, $1.50 against $0.30, and output tokens three times more, $7.50 against $2.50^[1]. For a developer that difference decides whether a task runs on every request or only when a cheaper model fails.

Gemini 3.5 Flash Cyber is the one release with a restricted audience. A model tuned to find vulnerabilities can be used to fix software before an attacker exploits it, or to look for weaknesses in systems its operator does not own; limiting it to governments and trusted partners in a pilot is a way of controlling the second use^[1].

Why it matters

The release shows three pressures operating on model makers at once. The first is cost: a cheaper, faster tier is what makes large-scale deployment — reading long documents, answering routine queries — economically sensible. The second is capability per token: 3.6 Flash is presented as better on coding and agent tasks while spending fewer output tokens, a claim about efficiency as much as about raw quality^[1]. The third is specialisation and controlled access: a security-focused model given to a narrow set of users rather than to everyone^[1].

The same announcement noted that Gemini 3.5 Pro was still in partner testing and that Gemini 4 pre-training had started, which places this launch inside a continuing cycle rather than at the end of one^[1]. For users of the Gemini app, the immediate change was that 3.6 Flash arrived worldwide on the day it was announced^[2].