Gemini 3.1 Pro endpoint tuned to pick your custom tools over bash in coding agents, with 1M-token context.
Output Speed *
Intelligence Index *
Context Window *
Input price
Output price
Performance
The Change Is Tool Selection; the Reasoning Base Is Gemini 3.1 Pro
Benchmarks
Independent Data Covers the Standard Endpoint: Strong Throughput, Slow First Token
Output Speed
Intelligence Index
MMLU *
GPQA *
HLE *
LiveCodeBench *
Technical Specifications
What the model supports
Comparison
Gemini 3.1 Pro Preview Custom Tools or Standard Gemini 3.1 Pro Preview for Agents?
These are two endpoints of the same Gemini 3.1 Pro preview. <cite index="1-3">Both are documented with a 1,048,576-token input limit and a 65,536-token output limit.</cite> <cite index="9-4">Vertex AI states that customtools pricing is identical to Gemini 3.1 Pro.</cite> The decision is about behavior: the standard endpoint is the general default, while customtools <cite index="9-2">better prioritizes custom tools</cite> in agents that also have bash. The question is whether your workload benefits from that bias.
Choose Gemini 3.1 Pro Preview Custom Tools when your agent exposes functions like view_file or search_code next to bash and the model keeps defaulting to shell commands. Choose standard Gemini 3.1 Pro Preview for general chat, analysis, writing, or coding without registered tools, where <cite index="1-6">Google warns customtools may show quality fluctuations.</cite> Run both on your own agent traces before committing, since no customtools-specific benchmark is published.
Limitations & Trade-offs
Best-Fit Workloads
Where this model earns its place
Coding agents with mixed tools
This endpoint targets agents that combine bash with registered functions. Google says it better prioritizes tools such as view_file and search_code. At launch, Artificial Analysis measured the standard endpoint at 54% on Terminal-Bench Hard and 59% on SciCode, both reported as top scores then. It fits repository navigation and edit-loop agents. Validate tool choice on your own traces, as customtools was not benchmarked separately.
Function-calling orchestration layers
When a planner model must choose among many internal APIs, over-reliance on a generic shell tool wastes steps and weakens control. This endpoint is tuned to favor user-defined functions, and Google lists function calling and structured outputs as supported on the model page. It suits support automation, internal-ops agents, and data-pipeline controllers. Keep deterministic validation on tool arguments.
Repository-scale code analysis
A 1,048,576-token context window and a 65,536-token output limit allow multi-file reviews and large refactor plans in one request. The agent can call search and file tools instead of loading everything. Retrieval reliability falls at extreme context length, so keep prompts well below the window for audits where a missed file matters.
Multimodal inputs feeding tool calls
The Gemini 3.1 Pro Preview page lists text, image, video, audio and PDF input. That lets an agent read a screenshot, recording or PDF and then act through registered tools, such as filing a ticket or querying a database. Output is text only, so media generation needs another model.