Developers and companies building production AI agents need three things above all: better token efficiency, lower latency, and reliable performance. To meet these demands, the Flash series strikes the sweet spot between efficiency and quality — and that balance is exactly what helps teams scale their agentic workflows effectively.
Takeaways 🚀
✅Faster models for AI agents
✅Lower latency, better efficiency
✅Reliable performance at scale
✅Flash models balance quality
✅Built for agentic workflows
With that goal in mind, we’re building on Gemini 3.5 Flash by introducing three new models today.
What’s New
- 3.6 Flash: Our workhorse model. It improves coding, knowledge work, and multimodal performance. On top of that, it uses 17% fewer output tokens than 3.5 Flash, according to the Artificial Analysis Index. In fact, some benchmarks — like DeepSWE by Datacurve — show savings of up to 65%. Best of all, it costs less per output token.
- 3.5 Flash-Lite: Our fastest and most cost-effective 3.5-class model. It delivers 350 output tokens per second and, notably, outperforms earlier Flash-Lite versions in agentic workflows.
- 3.5 Flash Cyber in CodeMender: A specialized, highly efficient model built for cybersecurity. We paired it with our CodeMender code security agent, and together, they reach competitive performance at the frontier.
That said, we’re not stopping there. Gemini 3.5 Pro is currently testing with partners, and we’ll release it broadly as soon as it’s ready. In parallel, our team has already started our most ambitious pre-training run yet: Gemini 4. We’re genuinely excited by the early progress.
3.6 Flash: More Efficient, Higher Quality
Gemini 3.6 Flash grew directly from developer and customer feedback on 3.5 Flash. As a result, it raises the bar on coding and knowledge work while also meaningfully improving token efficiency.
To put that in numbers: on the Artificial Analysis Index, 3.6 Flash uses 17% fewer output tokens than 3.5 Flash. Furthermore, it takes fewer reasoning steps and tool calls to complete multi-step workflows — making it noticeably leaner in practice.
What’s more, this efficiency gain comes with a lower price tag. At $1.50 per 1M input tokens and $7.50 per 1M output tokens, 3.6 Flash lowers the total cost per agentic task. In short, agents become cheaper and faster to build and run.
Stronger Performance Across the Board
Even with all these savings, 3.6 Flash still outperforms 3.5 Flash in several key areas:
- Coding: It makes fewer unwanted code edits and runs fewer execution loops. As a result, DeepSWE scores rose to 49% from 37%. ML research also improved sharply, with MLE Bench climbing to 63.9% versus 49.7%.
- Computer use: OSWorld-Verified scores climbed to 83.0% from 78.4%. Additionally, computer use is now a built-in, client-side tool through the Gemini API and Gemini Enterprise.
- Knowledge work: It outperforms 3.5 Flash on GDPval-AA v2, scoring 1421 versus 1349. Customers like Hebbia and Harvey have found it particularly strong at document parsing, chart and data analysis, and report drafting.
Taken together, customers describe 3.6 Flash as a clear step forward in both cost and quality — one that consistently balances token efficiency, accuracy, and speed across complex tasks.
Built With Safety
Beyond raw performance, 3.6 Flash also ships with stronger Frontier Safety safeguards. These cover Chemical, Biological, Radiological, and Nuclear (CBRN) risks, as well as cyber-offense misuse. As a result, the model is far more resistant to jailbreaks. At the same time, it was carefully trained to avoid unnecessary refusals for legitimate, beneficial uses — so it stays both safe and practical.
3.5 Flash-Lite: Built to Scale Agentic Workflows
Alongside 3.6 Flash, we’re also releasing Gemini 3.5 Flash-Lite. This model is designed specifically for low-latency tasks and high-throughput work — including agentic search and document processing.
3.5 Flash-Lite is the fastest model in the 3.5 series. According to Artificial Analysis, it runs at an impressive 350 output tokens per second. It’s also competitively priced at $0.30 per 1M input tokens and $2.50 per 1M output tokens. Combined with significantly better quality than 3.1 Flash-Lite, it offers a strong price-to-performance ratio for teams running high-volume production traffic.
Flexible Thinking Levels
One of 3.5 Flash-Lite’s standout features is how easily it scales agentic systems. Across all thinking levels, it clearly outperforms 3.1 Flash-Lite — giving developers much more flexibility to match the model to the task at hand.
Specifically, developers can choose minimal or low thinking levels for fast, low-cost, high-volume tasks. Alternatively, they can engage higher thinking levels to handle multi-step subagent workloads. In addition, the model includes computer use as a built-in tool, so it reliably supports agentic tasks across a wide range of surfaces.
These improvements show up clearly in benchmark results:
- Coding and agents: Terminal-Bench 2.1 rose to 54% from 31%.
- Long context: GDM-MRCR v2 climbed to 72.2% from 60.1%.
- Real-world tasks: GDPval-AA v2 jumped to 1140 from 642.
Moreover, on many agentic and coding tests, 3.5 Flash-Lite even outperforms 3 Flash. For instance, it scores 54.2% versus 49.6% on SWE-Bench Pro and 74.0% versus 65.1% on OSWorld-Verified. As a result, it becomes a faster, more capable option for a wide range of workloads.
Not surprisingly, early customers consistently highlight one standout quality: a rare combination of speed, intelligence, and cost efficiency that makes it ideal for scaling agentic workflows and data processing tasks.
3.5 Flash Cyber in CodeMender: Find and Fix Vulnerabilities Fast
As AI continues to advance, models are now finding security vulnerabilities faster than current systems can fix them. That growing gap calls for tools that are both highly capable and efficient — and that’s precisely where 3.5 Flash Cyber comes in.
Flash’s performance and efficiency make it an ideal foundation for security work. It can detect, validate, and patch code issues at scale. Building on that foundation, Gemini 3.5 Flash Cyber is fine-tuned specifically to find and fix cybersecurity vulnerabilities at a lower price per token than larger models.
Inside CodeMender, multiple 3.5 Flash Cyber agents work together to produce a single, combined report. On the widely used CyberGym benchmark, the model reaches competitive performance at the frontier — a strong result for a model built with efficiency in mind.
However, given the dual-use nature of this technology, we’re deploying it with care. For now, 3.5 Flash Cyber will be available exclusively to governments and trusted partners through CodeMender, as part of a limited-access pilot. This approach gives frontline defenders a meaningful head start on finding and fixing critical vulnerabilities — while also reducing the risk of broader misuse.
Get Started Today
Both 3.6 Flash and 3.5 Flash-Lite are available now across multiple platforms:
- Developers: Access them through the Gemini API via Google AI Studio and Android Studio. 3.6 Flash is also available in Google Antigravity. To get started right away, check the Developer Guide.
- Enterprises: Use them through the Gemini Enterprise Agent Platform. 3.6 Flash is additionally available in the Gemini Enterprise app.
- Everyone: Try them directly in the Gemini app. 3.5 Flash-Lite is also rolling out in Google Search.
We encourage you to start building with 3.6 Flash and 3.5 Flash-Lite today. As always, we welcome your feedback to help shape future Gemini models — and we look forward to
