Google just expanded its AI roster, but if you were bracing for a massive leap in raw reasoning, that’s not really the story here. Instead, the team rolled out Gemini 3.6 Flash, Gemini 3.5 Flash-Lite, and a security-tailored variant called Gemini 3.5 Flash Cyber.
Rather than competing for the top spot on benchmarks, this release zeroes in on the practical side of running AI systems, mainly trimming token usage, speeding up responses, and bringing down API bills for developers building agentic workflows.
Here is a closer look at what shipped and why it actually matters.
Gemini 3.6 Flash: Lower Bills Through Token Efficiency
The most interesting thing about Gemini 3.6 Flash isn’t just its lower headline price. According to benchmark runs from Artificial Analysis, the model uses around 17% fewer output tokens to complete identical tasks when compared to 3.5 Flash. Because it takes fewer reasoning steps and tool calls to get through complex prompts, the actual savings end up being much bigger than a basic price drop suggests.
On the technical side, Google bumped the pricing down to $1.50 per million input tokens and $7.50 per million output tokens (a noticeable drop from the old $9.00 output rate). They also refreshed the knowledge cutoff to March 2026, up from January 2025. It still supports a 1-million-token context window with a max output cap of 64k tokens, along with out-of-the-box features like client-side computer control, parallel tool calls, and structured outputs.
Benchmark performance saw solid bumps across the board:
- DeepSWE (Software Engineering): Climbed to 49%, up from 37% on the previous version.
- MLE Bench (Machine Learning): Jumped to 63.9%, compared to 49.7% previously.
- OSWorld-Verified (Computer Use): Reached 83.0%, up from 78.4%.
Gemini 3.5 Flash-Lite: Pure Speed for High-Volume Tasks
If you run high-frequency data pipelines or real-time sub-agents where waiting around for a response isn’t an option, Gemini 3.5 Flash-Lite is designed specifically for speed.
It cranks out throughput speeds up to 350 tokens every second. It’s also extremely cheap to run, coming in at $0.30 per million input tokens and $2.50 per million output tokens.
One useful detail for developers is the adjustable reasoning setting. You can manually dial the model’s “thinking” level anywhere from minimal to high, letting you save milliseconds on basic lookup jobs while giving the model more room to breathe on complex sub-tasks.
Gemini 3.5 Flash Cyber & CodeMender
Alongside the general-purpose models, Google showcased Gemini 3.5 Flash Cyber, a specialized model tuned specifically for security teams. It focuses on scanning code bases, identifying vulnerabilities, and automatically drafting fixes.
Because defensive tools like this carry obvious dual-use risks, Google isn’t releasing Flash Cyber as an open public API. Instead, they are integrating it directly into CodeMender, which is currently running in a limited pilot for vetted enterprise clients and government partners.
What About Gemini 3.5 Pro and Gemini 4?
Google also shared a quick update on where its larger frontier models stand:
- Gemini 3.5 Pro: Remains locked in private testing with enterprise partners. Google mentioned a broader public launch will happen once internal safety and quality benchmarks are met.
- Gemini 4: Pre-training for the next-generation foundation model has officially kicked off behind the scenes.
How to Try Them
Both Gemini 3.6 Flash and Gemini 3.5 Flash-Lite are live starting today. You can access them through the Gemini API (inside Google AI Studio and Android Studio), Gemini Enterprise, GitHub Copilot, the main Gemini app, and directly within Google Search experiences.
