# Gemini 3 Deep Think

> Google's extended parallel-reasoning mode on Gemini 3 Pro — built for hard math, science, and logic.

- Maker: Google
- Line: Gemini Deep Think
- Canonical page: https://ai-tldr.dev/models/gemini-3-deep-think/
- Updated: 2026-06-18

## In short

Gemini 3 Deep Think is Google's extended-reasoning mode, released December 4, 2025. It has a 1M-token context window. You can use it through Google AI Studio / Gemini API and Google Cloud Vertex AI.

## Facts

| Fact | Value | Source |
| --- | --- | --- |
| Context window | 1M (max output 64K tokens) | [1] |
| Input price | Included with Google AI Ultra | [2] |
| Output price | Included with Google AI Ultra | [2] |
| Licence | Proprietary · API only | [1] |
| Released | 3 December 2025 | [3] |
| Parameters | Undisclosed (sparse mixture-of-experts) | [1] |
| Knowledge cutoff | January 2025 | [1] |

## Sources

1. [Gemini 3 Pro Model Card (PDF) — Deep Think, specs, architecture](https://storage.googleapis.com/deepmind-media/Model-Cards/Gemini-3-Pro-Model-Card.pdf)
2. [Google AI Ultra plan (pricing)](https://blog.google/products-and-platforms/products/google-one/google-ai-ultra/)
3. [Gemini 3 Deep Think is now available in the Gemini app (Google blog)](https://blog.google/products/gemini/gemini-3-deep-think/)

## Overview

Gemini 3 Deep Think is Google's extended-reasoning mode, released December 4, 2025. It is not a separate model but an optional inference-time setting on Gemini 3 Pro: per Google's official Gemini 3 Pro model card, "Gemini 3 Pro now features Deep Think mode, an optional setting designed to enhance complex problem-solving performance at time of inference." When enabled, the model spends much more compute exploring several lines of thought in parallel before it commits to an answer, which is why a single Deep Think response can take minutes rather than seconds.

Because Deep Think runs on the Gemini 3 Pro base, it inherits that model's natively multimodal design — it accepts text, images, audio, and video, with a context window of up to 1 million tokens and output up to 64K tokens. The underlying architecture is a sparse mixture-of-experts (MoE) transformer with a January 2025 knowledge cutoff. Deep Think is aimed squarely at the hardest tasks: graduate-level science, competition mathematics, multi-step logic, and abstract visual reasoning.

At launch, Gemini 3 Deep Think rolled out to Google AI Ultra subscribers in the Gemini app (selectable as the "Deep Think" thinking mode), with API access offered through an early-access program for select researchers, engineers, and enterprises rather than as a separately-priced public endpoint. It posts standout reasoning scores: 41.0% on Humanity's Last Exam without tools, 93.8% on GPQA Diamond, and an ARC-Prize-verified 45.1% on ARC-AGI-2 with code execution.

## Benchmarks

- Humanity's Last Exam (no tools): 41% ([source](https://blog.google/products/gemini/gemini-3/))
- GPQA Diamond: 93.8% ([source](https://blog.google/products/gemini/gemini-3/))
- ARC-AGI-2 (with code execution, ARC Prize Verified): 45.1% ([source](https://blog.google/products/gemini/gemini-3/))

## Pricing

- Input: Included with Google AI Ultra subscription
- Output: Included with Google AI Ultra subscription

At launch Gemini 3 Deep Think had no separately-published per-token API price; consumer access was via the Google AI Ultra plan ($249.99/mo) in the Gemini app, with API access through an early-access program. Standard Gemini 3 Pro API rates apply to the base model.

Source: https://blog.google/products-and-platforms/products/google-one/google-ai-ultra/

## Strengths

- State-of-the-art abstract reasoning: 45.1% on ARC-AGI-2 (with code execution, ARC Prize Verified) — far above the standard Gemini 3 Pro result
- Top-tier graduate science knowledge: 93.8% on GPQA Diamond, ahead of Gemini 3 Pro's 91.9%
- Strong closed-book exam reasoning: 41.0% on Humanity's Last Exam with no tools
- Parallel-reasoning approach explores multiple hypotheses before answering, helping on hard math, physics, and logic problems
- Inherits the Gemini 3 Pro base: 1M-token context and native text/image/audio/video understanding
- Built for multi-step problem solving where extra inference-time compute pays off

## Best for

- Tackling competition-grade and research-level mathematics and proofs
- Graduate-level science questions in physics, chemistry, and related fields
- Abstract and visual reasoning puzzles that defeat standard chat models
- Reviewing technical or mathematical work to surface logical flaws
- Complex multi-step logic and planning problems
- Higher-quality code and prototypes for intricate, reasoning-heavy tasks

## How to access

- [Google AI Studio / Gemini API](https://ai.google.dev/gemini-api/docs/gemini-3) — model id `gemini-3-pro (Deep Think via thinking mode)`
- [Google Cloud Vertex AI](https://docs.cloud.google.com/vertex-ai/generative-ai/docs/models/gemini/3-pro) — model id `gemini-3-pro`

## Gemini Deep Think versions

- Gemini 3 Deep Think (3 December 2025) — this page
- [Gemini 2.5 Deep Think](https://ai-tldr.dev/models/gemini-2-5-deep-think/) (1 August 2025)

## FAQ

### Is Gemini 3 Deep Think a separate model from Gemini 3 Pro?

No. According to Google's official Gemini 3 Pro model card, Deep Think is an optional inference-time mode of Gemini 3 Pro, not a distinct model. Turning it on makes the model do more parallel reasoning before answering, so it shares Gemini 3 Pro's architecture, 1M-token context, and January 2025 knowledge cutoff.

### How do I access Gemini 3 Deep Think?

At its December 4, 2025 launch it was available to Google AI Ultra subscribers ($249.99/month) in the Gemini app, selected via the Deep Think thinking mode. API access was offered through an early-access program for select researchers, engineers, and enterprises rather than as a separately-priced public endpoint.

### How does Gemini 3 Deep Think score on benchmarks?

Per Google's launch blog, it reaches 41.0% on Humanity's Last Exam without tools, 93.8% on GPQA Diamond, and 45.1% on ARC-AGI-2 with code execution (ARC Prize Verified) — each above the standard Gemini 3 Pro result on the same tests.

### What is Gemini 3 Deep Think best at?

It targets the hardest reasoning tasks: competition-level mathematics, graduate science, multi-step logic, and abstract visual reasoning. The trade-off is latency — because it explores multiple hypotheses in parallel, a single answer can take minutes.

## Links

- [Gemini 3 Deep Think is now available in the Gemini app (Google blog)](https://blog.google/products/gemini/gemini-3-deep-think/)
- [Gemini 3 — official launch blog (benchmarks)](https://blog.google/products/gemini/gemini-3/)
- [Gemini 3 Pro Model Card (PDF) — Deep Think, specs, architecture](https://storage.googleapis.com/deepmind-media/Model-Cards/Gemini-3-Pro-Model-Card.pdf)
- [Gemini Deep Think — Google DeepMind](https://deepmind.google/models/gemini/deep-think/)
- [Gemini 3 Developer Guide (generateContent API)](https://ai.google.dev/gemini-api/docs/gemini-3)
- [Google AI Ultra plan (pricing)](https://blog.google/products-and-platforms/products/google-one/google-ai-ultra/)
- [Google rolling out Gemini 3 Deep Think to AI Ultra (9to5Google)](https://9to5google.com/2025/12/04/gemini-3-deep-think/)
- [Google Gemini 3 Benchmarks explained (Vellum)](https://www.vellum.ai/blog/google-gemini-3-benchmarks)

Source: AI/TLDR LLM registry, https://ai-tldr.dev/models/gemini-3-deep-think/