AI achieved perfect score in Math Olympiad

At the IMO exam in Shanghai earlier this week, RedNote said the company’s Dots-Note-3.0 model gave correct answers to all 6 questions, winning 42 points, equivalent to a gold medal, becoming the first AI to achieve this ability. Last year, two models from Google DeepMind and OpenAI also won gold medals, but solved 5/6 questions, reaching 35/42 points.

Theo China DailyDots-Note-3.0’s achievements show that AI has “advanced very far” in the field of reasoning and creative thinking. This year, only 7 contestants scored 42 points, accounting for 1%, while the standard score to win a gold medal is 29 points.

After RedNote’s announcement on July 22, AI expert Deedy Das of investment company Menlo Ventures (USA) also tested this year’s IMO test for four advanced artificial intelligence models, including Anthropic Claude Fable 5, OpenAI GPT 5.6 Sol, MoonShot Kimi K3 and Axiom Math. The results showed that all four successfully solved the entire test and won 42/42 points, of which Claude Fable 5 achieved the fastest speed.

“The frontier of AI has gone far beyond Math IMO,” Das wrote.

RedNote, known as Xiaohongshu in China, launched in 2013 in Shanghai, initially focusing on helping domestic users find and share overseas shopping experiences before evolving into a platform similar to a combination of Instagram, Pinterest and TikTok. According to SCMPthe platform currently has hundreds of millions of users, featuring travel, fashion, beauty, food and lifestyle content.

Unlike most older models that require humans to “translate” the problem into logical language for artificial intelligence to process, RedNote said the company’s AI operates directly on the original document without human intervention. The system operates from start to finish in accordance with the official rules of the competition, without receiving human intervention, such as giving suggestions, correcting reasoning steps or perfecting logic.

Dots-Note-3.0 is based on an “agent” framework that combines natural language reasoning with Python programming code execution to compose, test, and refine propositions when solving problems. On its website on July 22, the company explained that system performance comes from the ability to “self-check reasoning”, which includes questioning assumptions and continuously correcting logical errors through a multi-step self-evaluation loop before coming up with a final solution.

RedNote’s AI, currently in testing, is the lightest version of the company’s Dots3 line, along with Jazz and Aria, which are designed for private use and are both paid versions. On the website, the company plans to make the Dots-Note-3.0 source code public “in the near future”, but does not provide roadmap details.

IMO is one of the most difficult Math exams in the world. Originating in Romania in 1959, the contest is for contestants under 20 years old. The problems are selected from elementary mathematics, including the four areas of geometry, arithmetic, algebra and combinatorics. Unlike exams that are only based on answers and only need to evaluate the final result, problem solving in IMO requires evidence, requires rigorous reasoning, and must not have logical errors or missing conditions.

The 2026 International Mathematics Olympiad will be held in Shanghai from July 13 to 21, with the official exam date being July 15-16, attracting nearly 700 candidates from 119 countries and territories. This number is the largest in IMO’s 67-year history. The Vietnamese team had 6 students competing, winning two gold medals, three silver and one bronze, ranking 5th on the ranking of participating delegations.