# Gemini 2.5 Pro

Gemini 2.5 Pro is a Pro-tier model from Google, combining an enhanced base model with improved post-training to achieve a new level of reasoning performance in coding, mathematics, and scientific problem-solving.

- **Model ID:** `google/gemini-2.5-pro`
- **Type:** chat
- **Providers:** vertex, google
- **Context window:** 1,048,576
- **Maximum output tokens:** 65,535
- **Pricing:** $1.25/1M input tokens, $10/1M output tokens
- **Canonical page:** https://vercel.com/ai-gateway/models/gemini-2.5-pro

## Supported parameters

Detailed capability metadata has not been reported for this model.

## Example

```ts
import { streamText } from 'ai'

const result = streamText({
  model: 'google/gemini-2.5-pro',
  prompt: 'Why is the sky blue?'
})
```

## About

Google introduced Gemini 2.5 Pro on March 20, 2025 as the flagship of the Gemini 2.5 thinking model generation. Reasoning is its headline capability. Gemini 2.5 models reason through their thoughts before responding, and Google achieved this performance level by combining a significantly enhanced base model with improved post-training. On reasoning benchmarks, 2.5 Pro posts strong results on math and science (including GPQA and AIME 2025) without majority voting or other cost-increasing test-time techniques. On Humanity's Last Exam, a dataset designed by hundreds of subject matter experts to represent the human frontier of knowledge and reasoning, 2.5 Pro scores 18.8% without tool use.

Coding performance received particular attention. Gemini 2.5 Pro represents a significant leap over the 2.0 generation in creating web apps and agentic code applications, along with code transformation and editing. On SWE-Bench Verified, the industry-standard benchmark for agentic code evaluation, it scores 63.8% with a custom agent setup. It can generate a playable video game from a single-line prompt.

Gemini 2.5 Pro ships with a context window of 1.0M tokens, the largest among Gemini 2.5 models, and supports text, audio, images, video, and entire code repositories as input. Tool use including Google Search and code execution is available.

## What to consider

Given the context window of 1.0M tokens, applications passing very large inputs should confirm provider-side limits and latency expectations for long-context requests before deploying at scale.

## When to use

### Best For

- **Advanced coding and software engineering**, building visually compelling web applications, writing agentic code, performing large-scale code transformation and editing across entire repositories
- **Complex mathematical and scientific reasoning**, multi-step problems in mathematics, physics, chemistry, or logic that require sustained chain-of-thought reasoning without test-time augmentation
- **Research and long-document analysis**, processing entire codebases, academic papers, legal corpora, or research datasets within a context of 1.0M tokens to extract insights, connections, and answers
- **Hard benchmark-level tasks**, questions from expert-curated datasets, graduate-level reasoning problems, or tasks at the outer edge of what general-purpose models typically handle
- **Agentic applications requiring deep planning**, multi-step workflows where the model must reason across tools, plan sub-tasks, and produce executable or high-accuracy outputs

### Consider Alternatives When

- The task is high-volume and routine (translation, classification, summarization) where the reasoning depth of 2.5 Pro adds cost without improving output quality
- Response speed is paramount and accuracy requirements can be met by 2.5 Flash with thinking enabled
- Your application does not benefit from the context window of 1.0M tokens, making the pricing premium for Pro's larger capacity unnecessary
- The workload is primarily embedding or retrieval, where a dedicated embedding model is architecturally appropriate

## Best for

- **Advanced coding and software engineering:** building visually compelling web applications, writing agentic code, performing large-scale code transformation and editing across entire repositories
- **Complex mathematical and scientific reasoning:** multi-step problems in mathematics, physics, chemistry, or logic that require sustained chain-of-thought reasoning without test-time augmentation
- **Research and long-document analysis:** processing entire codebases, academic papers, legal corpora, or research datasets within a context of 1.0M tokens to extract insights, connections, and answers
- **Hard benchmark-level tasks:** questions from expert-curated datasets, graduate-level reasoning problems, or tasks at the outer edge of what general-purpose models typically handle
- **Agentic applications requiring deep planning:** multi-step workflows where the model must reason across tools, plan sub-tasks, and produce executable or high-accuracy outputs

## Consider alternatives

- **High-volume routine tasks:** Translation, classification, and summarization where the reasoning depth of 2.5 Pro adds cost without improving output quality
- **Speed-first accuracy targets:** Response speed is paramount and accuracy requirements can be met by 2.5 Flash with thinking enabled
- **Smaller context windows suffice:** Your application does not benefit from the context window of 1.0M tokens, making the pricing premium for Pro's larger capacity unnecessary
- **Embedding or retrieval workloads:** A dedicated embedding model is architecturally appropriate for these use cases

## Frequently asked questions

### What is Gemini 2.5 Pro's score on LMArena?

Gemini 2.5 Pro ranks highly on LMArena, which measures human preferences across a broad range of tasks. Check the LMArena leaderboard for the latest score, as rankings shift over time.

### What coding benchmarks does 2.5 Pro perform strongly on?

Gemini 2.5 Pro scores 63.8% on SWE-Bench Verified with a custom agent setup. SWE-Bench Verified is the industry-standard benchmark for agentic code evaluation. The model also excels at creating web apps, agentic code applications, and code transformation.

### How does 2.5 Pro's thinking capability differ from 2.5 Flash's?

Both models reason through problems before responding. Gemini 2.5 Pro is the Pro tier in the Gemini 2.5 family and posts strong results on coding, math, and science benchmarks. Gemini 2.5 Flash provides configurable thinking budgets and sits at the Pareto frontier of cost and performance.

### What is Humanity's Last Exam and how does Gemini 2.5 Pro perform on it?

Humanity's Last Exam is a benchmark dataset created by hundreds of subject matter experts to capture the human frontier of knowledge and reasoning. Gemini 2.5 Pro scores 18.8% on this benchmark without tool use.

### What is the context window size?

Gemini 2.5 Pro has a context window of 1.0M tokens, the largest among Gemini 2.5 models, enabling it to process entire code repositories, lengthy research datasets, or extensive multi-document inputs in a single pass.

### What tool use capabilities does 2.5 Pro have?

Google Search and code execution are available as built-in tools. The model can fetch real-time information, run code, and verify results within a single inference session.

### Does 2.5 Pro support multimodal input?

Yes. The model accepts text, audio, images, video, and entire code repositories as input, maintaining the native multimodality that defines the Gemini model family.

### Is Gemini 2.5 Pro generally available?

It launched as an experimental model on March 20, 2025. Google later promoted it to stable general availability as part of the Gemini 2.5 family expansion.

## Links

- [Model page](https://vercel.com/ai-gateway/models/gemini-2.5-pro)
- [AI Gateway documentation](https://vercel.com/docs/ai-gateway)
- [Provider model documentation](https://cloud.google.com/vertex-ai/generative-ai/docs/models/gemini/2-5-pro)
- [Provider pricing](https://cloud.google.com/vertex-ai/generative-ai/pricing)


If you encounter incorrect information, can’t find what you need, need a capability Vercel doesn’t currently offer, or have any other feedback to report, please do it by following the instructions at https://vercel.com/feedback.md.