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

Three new models in one day. Lower prices across the board. And one line most people scrolled past: 3.5 Flash-Lite is rolling out inside Google Search. Here is what shipped, what the benchmarks say, and why search teams should care.

Key Takeaways

• Google released three Gemini models on July 21, 2026: 3.6 Flash, 3.5 Flash-Lite and 3.5 Flash Cyber.
• Gemini 3.6 Flash uses 17% fewer output tokens than 3.5 Flash and costs less, at $1.50 per million input tokens and $7.50 per million output tokens.
• Gemini 3.5 Flash-Lite pushes 350 output tokens per second and is rolling out inside Google Search, where it handles agentic search.
• Gemini 3.5 Flash Cyber is a security model locked to governments and trusted partners through a CodeMender pilot.
• Gemini 3.5 Pro is in partner testing, and Google has started pre-training Gemini 4, which it calls its most ambitious run yet.
• Faster, cheaper models mean more AI answers per search. Being the cited source matters more than ever.

Three Models Dropped in a Single Post

Google did not stagger these launches. On July 21, product lead Tulsee Doshi pushed all three models live in a single announcement on The Keyword, with two teasers buried at the end: Gemini 3.5 Pro is in partner testing, and pre-training has started on Gemini 4.

The stated goal is blunt. Teams building production AI agents need cheaper tokens, lower latency and fewer failures. The Flash line is Google’s answer, and the headline numbers back it up.

Three Models, Three Very Different Jobs

Each model targets a distinct workload. One does the heavy thinking, one does the high-volume grunt work, and one is deliberately kept out of public hands.

The New Workhorse Does More While Writing Less

Gemini 3.6 Flash is the follow-up to the 3.5 Flash model Google shipped at I/O in May, and the pitch is unusual for an AI launch: it gets better by being quieter. Per the Artificial Analysis Index, 3.6 Flash consumes 17% fewer output tokens than 3.5 Flash. On Datacurve’s DeepSWE coding benchmark, Google observed cuts of up to 65%.

It also takes fewer reasoning steps and fewer tool calls to finish multi-step work. Stack that on a price cut, output pricing dropped from $9 to $7.50 per million tokens, and the cost of a finished agent task falls twice over.

Here is how the two models compare on the benchmarks Google published. OSWorld-Verified is the one to watch, since it measures how well a model drives a computer on its own.

Early production users include Figma, JetBrains, Harvey and Hebbia. The last two singled out document parsing, chart analysis and report drafting, the exact multimodal chores that eat analyst hours.

One more detail worth noting: 3.6 Flash ships with tighter Frontier Safety guardrails around chemical, biological, radiological, nuclear and cyber misuse. Google says it resists jailbreaks better than its predecessor while refusing fewer harmless requests.

Built for Speed, Priced for Volume

Flash-Lite is the model that runs all day. At 350 output tokens per second it is the fastest in the 3.5 series, and at $0.30 per million input tokens and $2.50 per million output tokens, it is priced for constant workloads: agentic search, document processing, product feed extraction.

Developers can dial thinking levels down for cheap, fast execution or up for multi-step subagent work, and computer use is built in. Against the 3.1 Flash-Lite it replaces, the jump is not subtle.

The more surprising result: on several agentic and coding tests, this Lite model beats the bigger Gemini 3 Flash outright.

The Sentence Everyone Scrolled Past: Flash-Lite Is Now Inside Google Search

Most coverage led with benchmarks. Barry Schwartz at Search Engine Land led with the line that matters for anyone who earns traffic from Google: 3.5 Flash-Lite is rolling out in Google Search, starting with agentic search. He also notes it could extend to AI Overviews and AI Mode, though Google has not spelled out the full footprint.

For context, Gemini 3.5 Flash has powered AI Mode globally since I/O in May. So the model behind Google’s AI answers has now been upgraded twice in about nine weeks. That cadence is the real story.

Each upgrade makes AI answers cheaper and faster to generate. Cheaper answers mean Google can afford to trigger them on more queries. Faster retrieval means the model reads, judges and quotes pages in milliseconds, and it favors pages it can lift from cleanly. Ecommerce sites building for AI buying agents are already living this shift.

What the New Models Cost Per Million Tokens

Token prices decide which workloads survive at scale. Both general-purpose models came in cheap, and one is not for sale at all.

The Security Model Stays Behind a Locked Door

Gemini 3.5 Flash Cyber is a fine-tuned version of 3.5 Flash built to find, validate and patch code vulnerabilities. Inside CodeMender, multiple Flash Cyber agents work in parallel and merge their findings into one report, and the setup posts frontier-competitive scores on the CyberGym benchmark.

Google is deliberately keeping this one off the open API. Access is limited to governments and trusted partners through a pilot program, the logic being that defenders should get a head start on patching before attackers get the same tooling.

Two more headlines hide at the bottom of the announcement. Gemini 3.5 Pro is testing with partners and lands when it is ready. And pre-training has begun on Gemini 4.

Five Moves for Agency Teams This Week

None of this requires new tools. It requires checking a few things before the next client call.

Quick Answers on the Gemini Release

AI Search Keeps Getting Faster. Keep Your Clients Cited.

Stan Ventures builds the third-party mentions, contextual links and citations that AI-powered search pulls from. White labeled, pre-approved domains, delivered in 25 days.

Book a Strategy Call