<MCP.Playground/>
online

Gemini 3.1 Pro

Runs MCP tools

Google’s frontier reasoning model, released in preview in February 2026. Google reports a 15-point jump on MCP Atlas over Gemini 3 Pro, a benchmark built on real MCP tool use.

Vendor
Google DeepMind
Released
19 February 2026
Context
1.05M tokens
Input
Text, image, video, audio, PDF

Paste a server URL, pick a model, and watch it call your tools in a real conversation. No install.

No server of your own? Leave it blank and use one of the public mock servers.

In your browser

MCP Agent Studio

Server has a public URL? Chat with Gemini 3.1 Pro on it, watch every tool call, and compare models side by side.

Open MCP Agent Studio →

On your computer

MCP Playground Desktop

Server runs on localhost or over stdio? The browser can’t reach it. The desktop app can, with Gemini 3.1 Pro in Agent Studio. Download free

Mac (Apple Silicon)Mac (Intel)Windowsor Microsoft Store

MCP Evals

Can Gemini 3.1 Pro actually complete tasks with your server?

One chat shows the tool calls work. An eval writes real tasks from your tool schemas, has Gemini 3.1 Pro drive each one, and reports where it got the wrong answer even though every call succeeded. Pick Gemini 3.1 Pro as the driver.

Run evals →See a sample

What Gemini 3.1 Pro is

Gemini 3.1 Pro is the Pro tier of Google’s Gemini 3 generation, built on the reasoning work behind Gemini 3 Deep Think. It is a mixture-of-experts model with a 1M-token window, up to 64K tokens of output, and input across text, images, video, audio and PDFs. The thinking_level parameter sets how long it reasons, and 3.1 adds a medium level between low and high. It is still labelled preview on OpenRouter, and Google has since shipped several Flash models, up to Gemini 3.8 Flash.

What Google DeepMind says

  • MCP Atlas: 69.2%, against 54.1% for Gemini 3 Pro.
  • Terminal-Bench 2.0: 68.5%, against 56.9% for Gemini 3 Pro.
  • ARC-AGI-2: 77.1%. GPQA Diamond: 94.3%.
  • BrowseComp: 85.9%.

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

  • Gemini CLI, Google’s open-source terminal agent, which connects to MCP servers
  • The Gemini API and Vertex AI, whose Google Gen AI SDKs can pass an MCP session to the model as tools
  • MCP Playground’s Agent Studio: paste your server URL above and run Gemini 3.1 Pro against it in the browser, with no Google key

What to watch for

  • Gemini 3 models return thought signatures with their function calls, and the API expects them back on the next request. A client that rebuilds the history from plain tool calls and drops the signatures can get errors or worse follow-up calls. On OpenRouter they travel in reasoning_details, so keep that field in the history too.
  • thinking_level changes how much it plans before the first tool call. It defaults to high. For simple lookups, low is enough; for chains where each step depends on the last, keep high. Medium is new in 3.1 and is worth testing first.
  • It is a preview model from February, and newer Flash models are cheaper and faster. Run the same task on Gemini 3.8 Flash before assuming you need Pro.

Run the same four-step task on Gemini 3.1 Pro and Gemini 3.8 Flash against the complex-schema mock: create_user_profile, process_order, analyze_data, then configure_workflow. If Flash makes the same calls, you do not need Pro for that server. All the mock servers →

Available here

ModelModel ID
Gemini 3.1 Progoogle/gemini-3.1-pro-preview

Go deeper

Frequently asked questions

Does Gemini 3.1 Pro support MCP?
Yes. Google reports 69.2% on MCP Atlas, a benchmark of real MCP tool use. Gemini CLI connects to MCP servers, and the Gen AI SDKs can pass an MCP session in as tools. You can run it against your own server here with no Google key.
What are thought signatures in Gemini 3?
Encrypted records of the model’s reasoning that come back with its function calls. Clients have to return them with the conversation so the model can pick up where it left off. Official Google SDKs handle this; hand-built history often loses them.
What thinking levels does Gemini 3.1 Pro have?
The thinking_level parameter takes low, medium and high, and defaults to high. Medium is new in 3.1. On OpenRouter it maps to reasoning_effort.
Gemini 3.1 Pro or Gemini 3.8 Flash for MCP?
Pro is the larger reasoning model; 3.8 Flash is newer, faster and cheaper per call. Both have 1M windows. Run the same prompt on each against your server here and compare the tool calls.

Sources