en flag
en flag

AI Won Gold at the Math Olympiad: What It Means

In 2025 an AI reached gold-medal level at the International Mathematical Olympiad, writing proofs in plain English. The facts, and what it means for you.

10/18/20263 min read

An AI Won Gold at the Math Olympiad. Here's What That Means, AI Prompt Agent featured graphic
An AI Won Gold at the Math Olympiad. Here's What That Means, AI Prompt Agent featured graphic

Every July, some of the best young mathematicians in the world sit down for two days of very hard problems at the International Mathematical Olympiad (IMO). The contest has been held almost every year since 1959. In 2025, an AI joined the conversation, and its results were officially graded.

Here's the short version of what happened, the facts behind it, and why it's more relevant to everyday AI users than it sounds.

New to prompts? Grab our free 10-prompt PDF for ChatGPT, Claude and Gemini.

What actually happened

On July 21, 2025, Google DeepMind announced that an advanced version of its Gemini model with "Deep Think" mode solved five of the six IMO 2025 problems perfectly, scoring 35 out of 42 points (Google DeepMind blog).

That matters because, according to the official IMO 2025 results, the gold medal threshold that year was 35 points (IMO official site). In other words, the AI landed exactly on the gold line.

The DeepMind post quotes IMO President Prof. Dr. Gregor Dolinar confirming the score and saying graders found the solutions "clear, precise and most of them easy to follow."

Why this was a big jump from 2024

The year before, DeepMind's AlphaProof and AlphaGeometry 2 systems reached silver-medal standard with 28 points. But there was a catch: human experts first had to translate the problems into a formal math language (Lean), and the computation took two to three days.

In 2025, the Gemini model worked end-to-end in natural language. It read the official problem text and wrote the proofs directly, within the 4.5-hour time limit the students get. That's the part that surprised many people: no special translation, just reading and reasoning.

The fine print (because it matters)

A few honest caveats, all from the same official sources:

  • It was a special version. DeepMind says the model was trained with extra reinforcement learning techniques and given a curated set of high-quality math solutions plus general tips for approaching IMO problems.

  • The IMO checked the answers, not the system. DeepMind notes that the IMO's review confirms the solutions were complete and correct, but doesn't validate the model or the process behind it.

  • The everyday version is lighter. When Deep Think rolled out in the Gemini app on August 1, 2025, Google said that version is faster for daily use and reaches bronze-level performance on IMO 2025 problems in internal tests (Google blog).

So no, your chatbot isn't secretly an Olympiad champion. But the gap between "research demo" and "app on your phone" is getting smaller every year.

What "thinking modes" mean for normal users

The interesting idea behind Deep Think is what DeepMind calls parallel thinking: instead of following one chain of reasoning, the model explores several possible solutions at the same time and combines them before answering.

You don't need to solve geometry problems to benefit from that. Reasoning modes, now offered in several AI tools, are useful whenever a task has more than one step:

  • comparing three pricing options for a product

  • planning a month of content around a launch

  • checking a contract summary for missing points

  • debugging a spreadsheet formula

How to prompt a "thinking" model better

Even the smartest model gives better answers when you ask well. Three habits borrowed from how math problems are framed:

1. State the goal and the constraints

Mathematicians never start without knowing exactly what must be proven. Do the same:

Goal: choose the best price for my [product]. Constraints: costs are [x], competitors charge [range], I want at least [margin]. Show your reasoning step by step, then give one recommendation.

2. Ask for alternatives before a final answer

This mimics the "explore several paths" idea:

Give me 3 different approaches to [problem], list the pros and cons of each, then recommend one and explain why.

3. Ask it to check its own work

Now review your answer as a strict critic. Point out any assumptions, missing information or possible errors, then give a corrected version.

A quick reality check

Benchmarks and competitions are impressive, but they're not the same as being right about your business, your health or your money. AI tools still make mistakes, sometimes confidently. Use them to think faster, and keep the final judgment yours.

Try it yourself

Want prompts that already use structures like "goal, constraints, alternatives, self-check"? Start with our free PDF, then explore the Pro Prompt Pack with 150 prompts for business, content and productivity.

Get better AI results today

Start with our free 10-prompt PDF, or get the full Pro Prompt Pack (150 prompts, $39).

Get the free 10 prompts · See the Pro Prompt Pack

FAQ

Did an AI really win a gold medal at the IMO?

AI systems don't receive official medals, since medals go to student contestants. Google DeepMind's model reached the gold-medal standard: 35 of 42 points, certified by IMO coordinators.

Can I use the same model?

Google rolled out a faster variant of Deep Think in the Gemini app for Google AI Ultra subscribers in August 2025. Google says it performs at bronze level on IMO 2025 problems in internal tests.

Does this mean AI is better than mathematicians?

No. Competition problems have known answers and fixed rules. Research mathematics is open-ended. The result shows strong reasoning on hard, well-defined problems.