# What is Gemini 2.5 Pro, and what makes it a "thinking" model?

Date: 2025-03-26 · Author: Robert Scales · https://guidetoartificialintelligence.com/en/learn/what-is-gemini-2-5-pro/

**Answer:** Gemini 2.5 Pro, released by Google on March 25, 2025, in an experimental version, is the first of a new Gemini generation in which every model reasons through a problem before answering. It debuts at the top of the LMArena leaderboard, leads several math and science tests, and can read about a million tokens, roughly 750,000 words, at once.

Yesterday, Tuesday, March 25, 2025, Google introduced **Gemini 2.5**, which it calls its most intelligent AI model so far. The first member of the family is **Gemini 2.5 Pro Experimental**, a test version of Google's most capable model, open to developers and to paying Gemini users from day one.

The headline feature is in the description Google gives it: a **thinking model**. Like OpenAI's o1 and o3-mini, DeepSeek's R1 and Anthropic's Claude 3.7 Sonnet, Gemini 2.5 Pro works through a problem step by step before writing its answer. Google says it will build this ability into all of its models from now on.

## What Google released

Koray Kavukcuoglu, chief technology officer of Google DeepMind, Google's AI research lab, announced the model on Google's blog. It is available in two places:

- **Google AI Studio**, Google's free website for developers to test its models.
- **The Gemini app**, on desktop and mobile, for subscribers to Gemini Advanced, the $20-a-month plan, who can pick it from the model menu.

Vertex AI, Google's cloud platform for businesses, gets it in the coming weeks. Prices for heavier professional use have not been announced yet; Google says they will follow soon.

This is not Google's first thinking model. In December it released an experimental "Flash Thinking" version of Gemini 2.0. What changed, according to Google, is the combination of a much stronger base model with better training after the main training phase, often called **post-training**.

## How a thinking model works

A regular language model writes its answer one word after another, straight away. A thinking model first produces a stretch of private reasoning: it breaks the question into parts, tries an approach, checks it and corrects itself. Only then does it write the reply you see.

Google traces the idea to two research threads it has pursued for years. **Chain-of-thought prompting** showed that models do better when asked to show their working. **Reinforcement learning** trains a model by rewarding good outcomes, which can teach it to reason more reliably.

Imagine planning a road trip with five stops, opening hours and a budget. A quick answer might list the stops in a random order. A thinking model is more likely to check distances, notice that one place is closed on Mondays and adjust the plan before showing it to you. The cost is time and computing: thinking takes longer and uses more resources.

## What the numbers show

Google and TechCrunch reported these results:

| Test | What it measures | Gemini 2.5 Pro |
|---|---|---|
| LMArena | Which answers people prefer in blind comparisons | Ranked first, by a clear margin |
| Humanity's Last Exam | Very hard expert questions, no tools allowed | 18.8% |
| SWE-bench Verified | Fixing real software bugs, with a custom agent setup | 63.8% |
| Aider Polyglot | Editing code across programming languages | 68.6% |

Google also says the model leads on GPQA (graduate-level science) and AIME 2025 (competition math) without costly tricks such as majority voting, where a model answers many times and the most common answer wins.

Scores need context. **Humanity's Last Exam** was written by hundreds of specialists to be extremely hard, so 18.8% is a strong result among models that do not use tools. On SWE-bench Verified, TechCrunch notes, Gemini 2.5 Pro beats o3-mini and R1 but trails Claude 3.7 Sonnet's 70.3%. No model wins everything.

## The long memory

Gemini models have been known for their large **context window**, the amount of material the model can consider in one request. Gemini 2.5 Pro ships with one million tokens, and Google says two million is coming soon. A token is a piece of a word; TechCrunch puts a million tokens at about 750,000 words, more than the whole *Lord of the Rings* series.

The model is also **multimodal**: it can work with text, audio, images, video and even an entire collection of code files. Google stresses coding in particular, showing the model writing a working video game from a one-line request and building polished web pages.

## Why it matters

Since OpenAI introduced o1 last September, reasoning has become the main way AI labs push their models forward, especially in math, science and coding. Gemini 2.5 Pro is Google's strongest move yet in that race, and the top LMArena ranking suggests people also like how it writes, not only how it scores.

The other notable point is direction. Google says thinking will no longer be a special edition but a standard part of every Gemini model. For users, that means choosing a model may matter less than learning how to give a thinking model a well-defined problem.

> **Try it:** Give Gemini 2.5 Pro a planning problem with real constraints, such as a week of family dinners for a budget of $120 with one vegetarian and no nuts. Read its answer and check two things yourself: whether the total really fits the budget, and whether every rule was respected.

## How can you try Gemini 2.5 Pro today?

If you are a developer or simply curious, go to Google AI Studio and sign in with a Google account. Select Gemini 2.5 Pro Experimental from the model list and start typing. It is a good place to test long inputs: paste in a long report or upload several files and ask questions that require connecting information from different parts.

If you subscribe to Gemini Advanced, open the Gemini app on your computer or phone and choose 2.5 Pro from the model menu. For now, the app version is limited to Advanced subscribers.

Because it is experimental, expect changes and limits on how much you can use it. Give it clear, complete problems rather than one-word questions, state any rules it must follow, and check its work on anything that matters, especially numbers. A useful starting prompt:

```prompt
Here is my problem: [describe it fully].
List the constraints you will respect, solve it step by step,
then give me the final answer in a short table and point out anything you were unsure about.
```

## Sources

- Gemini 2.5: Our most intelligent AI model: https://blog.google/technology/google-deepmind/gemini-model-thinking-updates-march-2025/
- Google unveils a next-gen family of AI reasoning models: https://techcrunch.com/2025/03/25/google-unveils-a-next-gen-ai-reasoning-model/