Gemini 3.8 Flash
Runs MCP tools
Google’s newest Flash model, released 2 September 2026 and its third Flash release in six weeks. Google calls it its best reasoning and coding model yet, at the speed of 3.7 Flash.
What Gemini 3.8 Flash is
Gemini 3.8 Flash succeeds 3.7 Flash. It has a 1M-token window, up to 64K tokens of output, and takes text, image, video, audio and file input. Google released it alongside 3.8 Flash Cyber, a separate variant for vulnerability detection and patching that is limited to vetted defenders and not available here.
What Google says
- Terminal-bench 2.1: 89.4%.
- DeepSWE v1.1: 73.7%.
- Terminal-bench 4.0: 19.1%.
- GDPval-AA v2: 1545, up from 1482 for 3.7 Flash.
- A significant gain in prompt-injection robustness on Gray Swan’s benchmark, by Google’s account.
These are the vendor’s own claims, not measurements of ours. Run the model against your own server to find out whether they hold for your tools.
How it reaches MCP
- The Gemini API, whose Python and JavaScript SDKs can take an MCP session as a tool directly. Google marks this experimental.
- MCP Playground’s Agent Studio: paste your server URL above and run Gemini 3.8 Flash against it in the browser, with no Google key
- Gemini CLI, Google’s terminal agent, which acts as an MCP client
What to watch for
- Google’s own numbers split sharply: 89.4% on Terminal-bench 2.1 and 19.1% on Terminal-bench 4.0. Short tool tasks may go well and long multi-step chains may not. Test the length of chain your server actually needs.
- MCP tool results are a prompt-injection route, and Google says 3.8 Flash is significantly more robust on Gray Swan’s benchmark. That is worth checking against your own server’s outputs rather than assuming.
- Three Flash releases in six weeks means the model behind a “latest” alias moves quickly. Pin gemini-3.8-flash if you need results to stay comparable.
Run 3.8 Flash and 3.7 Flash on the same four-step task against the complex-schema mock: create_user_profile, process_order, analyze_data, then configure_workflow. The Terminal-bench split suggests the later steps are where the two will differ. All the mock servers →
Available here
| Model | Model ID |
|---|
| Gemini 3.8 Flash | google/gemini-3.8-flash |
Frequently asked questions
Does Gemini 3.8 Flash support MCP?
Yes. Google’s Gemini SDKs for Python and JavaScript can use an MCP session as a tool directly, marked experimental, and any MCP client can drive it through function calling. You can run it against your own server here without a Google key.
What is Gemini 3.8 Flash Cyber?
A separate variant built for finding and patching software vulnerabilities, released the same day. Access is limited to vetted defenders through Google’s application programme, and it is not available here.
What changed from Gemini 3.7 Flash?
Google reports gains across software engineering, agent tasks and multi-step reasoning at the same speed, including GDPval-AA v2 up from 1482 to 1545. Both are available here, so you can compare them on your tools.
Can Gemini 3.8 Flash read audio and video from MCP tools?
It accepts audio and video input, but MCP tool results come back as text, image or embedded resources. Whether audio or video reaches the model depends on your client passing it through, so test with text and image results first.