# Gemini 2.5 Deep Think

> Gemini 2.5 Pro's parallel-thinking reasoning mode, with an IMO-medal math variant.

- Maker: Google
- Line: Gemini Deep Think
- Canonical page: https://ai-tldr.dev/models/gemini-2-5-deep-think/
- Updated: 2026-06-18

## In short

Gemini 2.5 Deep Think is Google's enhanced reasoning mode for Gemini 2.5 Pro, which became generally available on 1 August 2025 inside the Gemini app for Google AI Ultra subscribers. It has a 1M-token context window. You can use it through Google Gemini API (trusted testers).

## Facts

| Fact | Value | Source |
| --- | --- | --- |
| Context window | 1M (max output 192K) | [1] |
| Licence | Proprietary · API only | [1] |
| Released | 1 August 2025 | [2] |
| Parameters | Undisclosed | [1] |
| Knowledge cutoff | April 2024 | [1] |

## Sources

1. [Gemini 2.5 Deep Think Model Card (Google DeepMind, PDF)](https://storage.googleapis.com/deepmind-media/Model-Cards/Gemini-2-5-Deep-Think-Model-Card.pdf)
2. [Gemini 2.5: Deep Think is now rolling out (Google blog)](https://blog.google/products-and-platforms/products/gemini/gemini-2-5-deep-think/)

## Overview

Gemini 2.5 Deep Think is Google's enhanced reasoning mode for Gemini 2.5 Pro, which became generally available on 1 August 2025 inside the Gemini app for Google AI Ultra subscribers. Rather than a separate base model, Deep Think is a way of running Gemini 2.5 Pro that lets it think for much longer and explore many lines of reasoning in parallel before it answers.

Its core idea is 'parallel thinking'. Where a normal model follows one chain of thought, Gemini 2.5 Deep Think generates several candidate solutions at the same time, then revises and combines them over time to reach a stronger final answer. Google built it on a sparse mixture-of-experts transformer, accepts text and image input, and gives it a 1M-token context window with up to roughly 192K tokens of output.

Google positioned Gemini 2.5 Deep Think for the hardest math, science, and competition-coding problems. It posted state-of-the-art results on Humanity's Last Exam and LiveCodeBench v6 at launch, and a related variant reached gold-medal standard at the 2025 International Mathematical Olympiad. It has since been superseded by Gemini 3 Deep Think.

## Benchmarks

- Humanity's Last Exam (no tools): 34.8% ([source](https://techcrunch.com/2025/08/01/google-rolls-out-gemini-deep-think-ai-a-reasoning-model-that-tests-multiple-ideas-in-parallel/))
- LiveCodeBench v6: 87.6% ([source](https://techcrunch.com/2025/08/01/google-rolls-out-gemini-deep-think-ai-a-reasoning-model-that-tests-multiple-ideas-in-parallel/))
- MMMU (multimodal reasoning): 84% ([source](https://blog.google/innovation-and-ai/models-and-research/google-deepmind/google-gemini-updates-io-2025/))
- AIME 2025: 99.2% ([source](https://www.techtarget.com/whatis/feature/Gemini-25-Deep-Think-explained))
- IMO 2025 benchmark (Bronze): 60.7% ([source](https://www.techtarget.com/whatis/feature/Gemini-25-Deep-Think-explained))

## Strengths

- Parallel thinking explores and combines multiple solution paths, which lifts performance on genuinely hard, multi-step problems.
- State-of-the-art reasoning at launch: 34.8% on Humanity's Last Exam without tools, ahead of Grok 4 and OpenAI o3.
- Strong competition-level coding: 87.6% on LiveCodeBench v6.
- Elite math, including 99.2% on AIME 2025 and Bronze-level (60.7%) on the 2025 IMO benchmark.
- Large 1M-token context window for working over long documents and codebases.

## Best for

- Hard mathematics and proofs, including olympiad-style and research-level problems.
- Competition-grade and complex algorithmic coding tasks.
- Graduate- and expert-level science reasoning where a single chain of thought is not enough.
- Multi-step analysis that benefits from weighing several hypotheses before answering.
- Long-context problem solving across large documents thanks to the 1M-token window.

## How to access

- [Google Gemini API (trusted testers)](https://blog.google/products-and-platforms/products/gemini/gemini-2-5-deep-think/) — model id `gemini-2.5-deep-think`

## Gemini Deep Think versions

- [Gemini 3 Deep Think](https://ai-tldr.dev/models/gemini-3-deep-think/) (3 December 2025)
- Gemini 2.5 Deep Think (1 August 2025) — this page

## FAQ

### What is Gemini 2.5 Deep Think?

It is Google's enhanced reasoning mode for Gemini 2.5 Pro, released on 1 August 2025. It uses 'parallel thinking' to explore many ideas at once and combine them before answering, which improves results on hard math, science, and coding problems.

### How is Deep Think different from regular Gemini 2.5 Pro?

Deep Think runs the same Gemini 2.5 Pro model but lets it think far longer and reason over multiple solution paths in parallel rather than a single chain of thought. That extra compute is what drives its higher benchmark scores, at the cost of slower responses.

### How good is Gemini 2.5 Deep Think on benchmarks?

At launch it was state of the art on several hard tests: 34.8% on Humanity's Last Exam without tools (ahead of Grok 4 and OpenAI o3), 87.6% on LiveCodeBench v6, 99.2% on AIME 2025, and Bronze-level (60.7%) on the 2025 IMO benchmark.

### How much does Gemini 2.5 Deep Think cost?

There was no public per-token API price at launch. Consumers got it through the Google AI Ultra subscription at $249.99/month in the US, with a fixed number of prompts per day; broader Gemini API access was limited to trusted testers.

## Links

- [Gemini 2.5: Deep Think is now rolling out (Google blog)](https://blog.google/products-and-platforms/products/gemini/gemini-2-5-deep-think/)
- [Gemini 2.5 Deep Think Model Card (Google DeepMind, PDF)](https://storage.googleapis.com/deepmind-media/Model-Cards/Gemini-2-5-Deep-Think-Model-Card.pdf)
- [Google I/O 2025: Updates to Gemini 2.5 (Google blog)](https://blog.google/innovation-and-ai/models-and-research/google-deepmind/google-gemini-updates-io-2025/)
- [Google rolls out Gemini Deep Think (TechCrunch)](https://techcrunch.com/2025/08/01/google-rolls-out-gemini-deep-think-ai-a-reasoning-model-that-tests-multiple-ideas-in-parallel/)
- [Gemini 2.5 Deep Think explained (TechTarget)](https://www.techtarget.com/whatis/feature/Gemini-25-Deep-Think-explained)
- [Gemini models — Google DeepMind](https://deepmind.google/models/gemini/)

Source: AI/TLDR LLM registry, https://ai-tldr.dev/models/gemini-2-5-deep-think/