Introducing Gemini 3.6 Flash, 3.5 Flash-Lite, and 3.5 Flash Cyber

What it claims

Google DeepMind released two generally available models and announced one gated model. Gemini 3.6 Flash improves on 3.5 Flash in coding, knowledge work and computer use. It also uses fewer output tokens, reasoning steps and tool calls per task, and costs less per output token. Gemini 3.5 Flash-Lite is a high-throughput tier that beats the older 3 Flash on several agentic benchmarks. The page positions it as a replacement for 2.5 and 3 Flash workloads. Gemini 3.5 Flash Cyber is 3.5 Flash fine-tuned for finding and fixing vulnerabilities. It runs only inside the CodeMender agent and is not yet available.

Evidence

All figures are as stated by Google. Token-efficiency and speed figures come from Artificial Analysis.

3.6 Flash vs 3.5 Flash

• Output-token reduction is 17% on the Artificial Analysis Index and up to 65% on DeepSWE.
• On DeepSWE the page credits the gain to fewer unwanted code edits and fewer execution loops.

3.5 Flash-Lite vs 3.1 Flash-Lite

3.5 Flash-Lite vs 3 Flash

3.5 Flash Cyber: the page says only that it reaches "competitive performance at the frontier" on CyberGym. It gives no number and no comparison model.

How to use it

Pricing per 1M tokens

Thinking levels (Flash-Lite): the model takes configurable levels.

• Use the minimal or low levels for low-latency, low-cost bulk work such as agentic search and document processing.
• Use higher levels for multi-step subagent workloads.

Computer use is now a built-in client-side tool in the Gemini API and Gemini Enterprise, for both models. Because it runs client-side, your client executes the actions.

Availability, starting now:

• Gemini API via Google AI Studio and Android Studio. 3.6 Flash is also in Google Antigravity.
• Gemini Enterprise Agent Platform. 3.6 Flash is also in the Gemini Enterprise app.
• The Gemini app. Flash-Lite is also rolling out in Google Search.

Further reading: the page points to separate model cards for both models and to a Developer Guide.

Limits

• All results are self-reported. The page gives no evaluation settings, such as thinking level, scaffold or number of runs.
• Flash-Lite's configuration is unstated. Its benchmarks are not tied to a thinking level, even though the page says quality varies across levels.
• The cost comparison is incomplete. The page says 3.6 Flash is cheaper than 3.5 Flash but does not give 3.5 Flash's price.
• 3.5 Flash Cyber cannot be used yet. It will "soon" be available, exclusively to governments and trusted partners, through a limited-access CodeMender pilot. Inside CodeMender, multiple Flash Cyber agents work together and their findings are merged into one report.
• 3.6 Flash has stronger safeguards against CBRN and cyber-offense misuse. Google says the model is substantially more resistant to jailbreaks and was trained to minimise refusals for beneficial uses. If your security tooling sits near these domains, test it for refusals.
• Upcoming models are not available. Gemini 3.5 Pro is still in partner testing with no date, and Gemini 4 is in pre-training.