# Google's new Flash models chase the same goal: cheaper, faster AI agents > Gemini 3.6 Flash, 3.5 Flash-Lite, and a security-only 3.5 Flash Cyber all aim at running AI agents at scale — for less money, with fewer wasted words. _Source: Google (first-party announcement) · 2026-07-21 · 5 min read · Verified against primary sources_ Canonical: https://iyu.app/e/gemini-3-6-flash-family ## The 60-second version Google launched three new Gemini Flash models — 3.6 Flash, 3.5 Flash-Lite, and a security-only 3.5 Flash Cyber — all tuned to run AI agents cheaply and fast, with efficiency (fewer tokens per task) as the pitch. **Key points** - 3.6 Flash: the workhorse — Google claims ~17% fewer output tokens than 3.5 Flash, plus a price cut to $1.50/$7.50 per 1M tokens. - 3.5 Flash-Lite: fastest and cheapest ($0.30 input), 350 tokens/sec; Google says it beats the older, bigger '3 Flash' on some agentic tests. - 3.5 Flash Cyber: a dual-use vulnerability finder/fixer, held back to governments and trusted partners via the CodeMender pilot. - 3.5 Pro is coming soon; pre-training on Gemini 4 has started. **Verdict.** A clear bet that the agent era rewards cheap efficiency over raw size — but nearly all the numbers are Google's own, so treat them as promising claims, not proven results. ## Full explainer > **⚑ Caveat:** Nearly every figure below — token savings, benchmark scores, speed — is **Google's own**, or drawn from indices Google chose to cite. Treat them as vendor claims until independent testing confirms them. ### What shipped — Three models, one obsession: efficiency Google refreshed its cheap-and-fast **Gemini Flash** line with three new models, all pointed at the same target: running AI **agents** — systems that grind through multi-step tasks by calling tools and checking their own work — affordably and at scale. The common thread isn't a bigger brain; it's doing the same work with fewer tokens, which is the thing agents actually cost money on. The lineup: **3.6 Flash**, a stronger everyday workhorse; **3.5 Flash-Lite**, an ultra-cheap speed model; and **3.5 Flash Cyber**, a security-only variant locked behind a pilot program. ### The workhorse — 3.6 Flash: same job, fewer words The headline for 3.6 Flash isn't a benchmark — it's frugality. Google says it uses about **17% fewer output tokens** than 3.5 Flash on the third-party Artificial Analysis Index (and up to 65% fewer on some coding tests), while taking fewer reasoning steps and tool calls to finish a task. It also got cheaper. Since agents pay per token, fewer tokens per task means real savings. - **17%** — fewer output tokens vs 3.5 Flash (Artificial Analysis Index) - **$1.50** — per 1M input tokens - **$7.50** — per 1M output tokens On quality, Google reports gains across the board versus 3.5 Flash — more precise coding, a large jump in machine-learning research tasks, and better **computer use** (the model driving a screen the way a person would), now a built-in tool in the Gemini API. - **Benchmark (3.6 vs 3.5 Flash):** Score - **DeepSWE (coding):** 49% vs 37% - **MLE Bench (ML research):** 63.9% vs 49.7% - **OSWorld-Verified (computer use):** 83.0% vs 78.4% - **GDPval-AA v2 (knowledge work):** 1421 vs 1349 *Figure: Google's own quality-benchmark chart for 3.6 Flash.* ### The budget speedster — 3.5 Flash-Lite: cheap, fast, surprisingly strong Flash-Lite is the fastest model in the 3.5 family — Google clocks it at **350 output tokens per second** — and the cheapest, at $0.30 per 1M input tokens and $2.50 output. Its job is high-volume, low-latency work: agentic search, document processing, big pipelines. Developers can dial its 'thinking' up or down to trade speed for depth. The eye-catching claim: on several agentic and coding tests, Google says Flash-Lite **beats the older, larger '3 Flash'** — e.g. SWE-Bench Pro (54.2% vs 49.6%) and OSWorld-Verified (74.0% vs 65.1%). Smaller and cheaper, but not necessarily weaker. ### The locked door — 3.5 Flash Cyber: a security specialist, held back on purpose The third model is different in kind. **3.5 Flash Cyber** is a version of Flash fine-tuned purely to find and fix software vulnerabilities, running inside a system called **CodeMender**. The design twist: CodeMender runs several Flash Cyber agents together and merges their work into a single combined report. Google says it reaches competitive frontier performance on the CyberGym benchmark. > **🔒** Google is explicit that this is **dual-use** — the same skill that patches a bug can find one to exploit. So Flash Cyber won't be broadly released: it's limited to **governments and trusted partners** via a CodeMender pilot, to give defenders a head start while limiting misuse. ### Why it matters — The industry is pivoting from 'biggest' to 'cheapest that works' Zoom out and the launch tells a bigger story: the race is shifting from *how smart can one giant model be* to *how cheaply can a good-enough model run millions of agent tasks*. Efficiency, not raw scale, is the new battleground — because agents multiply every token by thousands of steps. > For AI agents, the winning model isn't the smartest — it's the one that reaches the answer using the fewest tokens. Google added that **Gemini 3.5 Pro** is testing with partners and coming soon, and — the real eyebrow-raiser — that pre-training has begun on **Gemini 4**. > **⚡** 3.6 Flash and 3.5 Flash-Lite are available today via the Gemini API (Google AI Studio, Android Studio), Gemini Enterprise, and the Gemini app; Flash-Lite is also rolling out in Google Search. ## Primary sources - [Google — 3.6 Flash, 3.5 Flash-Lite, and 3.5 Flash Cyber](https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-3-6-flash-3-5-flash-lite-3-5-flash-cyber/) --- _Published by iyu (https://iyu.app) — the day's AI news, checked against primary sources and rewritten in plain language. Free to quote with attribution and a link to the canonical URL._