Just yesterday, the world's top large models were completely defeated at the IMO competition in 2025, not even touching the edge of a bronze medal.
However, just now, OpenAI dropped a heavy bombshell—they successfully won the gold medal at IMO 2025 with a brand new &#34;general reasoning model&#34;!
Out of 6 problems, they solved 5, racking up an impressive 35 points!
It's worth noting that the previous best performer, Gemini 2.5 Pro, only scored 13 points.
<img src="https://imageproxy.pbkrs.com/https://bang.xuangubao.cn/MTQ1MDA1MTQ5MQ==?x-oss-process=image/auto-orient,1/interlace,1/resize,w_1440,h_1440/quality,q_95/format,jpg" alt="" original-src="https://imageproxy.pbkrs.com/https://bang.xuangubao.cn/MTQ1MDA1MTQ5MQ=="/>
Co-founder Greg Brockman, lead Alexander Wei, and various researchers from OpenAI excitedly announced this milestone achievement on Twitter!
<img src="https://imageproxy.pbkrs.com/https://bang.xuangubao.cn/LTgyNDM4MjIwNA==?x-oss-process=image/auto-orient,1/interlace,1/resize,w_1440,h_1440/quality,q_95/format,jpg" alt="" original-src="https://imageproxy.pbkrs.com/https://bang.xuangubao.cn/LTgyNDM4MjIwNA=="/>
In response, Noam Brown, the father of poker, stated that the significance of this achievement goes beyond &#34;AI conquering the IMO&#34; itself.
<img src="https://imageproxy.pbkrs.com/https://bang.xuangubao.cn/LTE0Mjc2OTUxMDc=?x-oss-process=image/auto-orient,1/interlace,1/resize,w_1440,h_1440/quality,q_95/format,jpg" alt="" original-src="https://imageproxy.pbkrs.com/https://bang.xuangubao.cn/LTE0Mjc2OTUxMDc="/><img src="https://imageproxy.pbkrs.com/https://bang.xuangubao.cn/MTQzNTY3NjE4OA==?x-oss-process=image/auto-orient,1/interlace,1/resize,w_1440,h_1440/quality,q_95/format,jpg" alt="" original-src="https://imageproxy.pbkrs.com/https://bang.xuangubao.cn/MTQzNTY3NjE4OA=="/><img src="https://imageproxy.pbkrs.com/https://bang.xuangubao.cn/LTEzODU5Mzk3MTI=?x-oss-process=image/auto-orient,1/interlace,1/resize,w_1440,h_1440/quality,q_95/format,jpg" alt="" original-src="https://imageproxy.pbkrs.com/https://bang.xuangubao.cn/LTEzODU5Mzk3MTI="/>
Swipe left and right to view
People are speculating that OpenAI may have deployed a disruptive reasoning technology this time, completely bidding farewell to the traditional Chain of Thought (CoT) reasoning.
This is not just a victory for a model, but the beginning of a whole new era!
Even more shocking is that OpenAI announced that this historic model is not the rumored GPT-5, but a completely new, experimental model!
And this model will not be released at all, it's simply too mysterious!
Mysterious model wins IMO gold medal
Alexander Wei and his team had the model compete under the exact same conditions as humans:
Two 4.5-hour exam periods, no tools, no internet, relying solely on understanding the problems to write out complete proofs in natural language.
Then, it was strictly graded by three former IMO medalists.
In the end, the model achieved an astonishing high score of 35/42, reaching gold medal level.
<img src="https://imageproxy.pbkrs.com/https://bang.xuangubao.cn/LTE5NjI3NDIwNDE=?x-oss-process=image/auto-orient,1/interlace,1/resize,w_1440,h_1440/quality,q_95/format,jpg" alt="" original-src="https://imageproxy.pbkrs.com/https://bang.xuangubao.cn/LTE5NjI3NDIwNDE="/><img src="https://imageproxy.pbkrs.com/https://bang.xuangubao.cn/ODc0MDE3NjEw?x-oss-process=image/auto-orient,1/interlace,1/resize,w_1440,h_1440/quality,q_95/format,jpg" alt="" original-src="https://imageproxy.pbkrs.com/https://bang.xuangubao.cn/ODc0MDE3NjEw"/>
In contrast, in the past, AI needed specialized training in specific fields to achieve victory in complex tasks like Go, Dota, or others
But this time, OpenAI broke this iron rule—the new model not only is not a special supply for IMO, but can also think for hours.
In contrast, the o1 model we are familiar with is measured in seconds, and Deep Research is measured in minutes.
This deep, sustained creative thinking ability is a chasm that AI has struggled to overcome in the past!
<img src="https://imageproxy.pbkrs.com/https://bang.xuangubao.cn/MTYzMDM3NjkxMw==?x-oss-process=image/auto-orient,1/interlace,1/resize,w_1440,h_1440/quality,q_95/format,jpg" alt="" original-src="https://imageproxy.pbkrs.com/https://bang.xuangubao.cn/MTYzMDM3NjkxMw=="/>
What does this mean? Has it reached AGI level? What is special about winning IMO?
First, compared to previous benchmarks, IMO problems require a higher level of sustained creative thinking.
In terms of reasoning time, it has now climbed all the way: GSM8K (top humans need about 0.1 minutes) → MATH benchmark (about 1 minute) → AIME (about 10 minutes) → IMO (about 100 minutes).
Secondly, the submissions for IMO are difficult to verify multi-page proofs.
Making progress in this field requires surpassing the reinforcement learning paradigm with clear, verifiable rewards.
By doing so, the OpenAI research team has developed a model capable of constructing complex and impeccable arguments at the level of human mathematicians.
Moreover, this model did not achieve this level of capability through training on &#34;specific tasks (IMO),&#34; but rather achieved new breakthroughs in general reinforcement learning and computational expansion during testing.
<img src="https://imageproxy.pbkrs.com/https://bang.xuangubao.cn/NzY1NzQwMzk3?x-oss-process=image/auto-orient,1/interlace,1/resize,w_1440,h_1440/quality,q_95/format,jpg" alt="" original-src="https://imageproxy.pbkrs.com/https://bang.xuangubao.cn/NzY1NzQwMzk3"/>
So, was the o3-alpha exposed last night just a prelude?
It turns out this is the real trump card that OpenAI has been holding back!
Complete Solution Process
If you are interested, you can take a look at the freshly released 2025 IMO competition problem solutions from OpenAI.
The model solved problems one to five (P1-P5), but failed to solve problem six (P6). As rumored, the difficulty of this sixth problem is extremely high, with only 6 people worldwide cracking it.
<img src="https://imageproxy.pbkrs.com/https://bang.xuangubao.cn/MTYxNDk1MzEzOQ==?x-oss-process=image/auto-orient,1/interlace,1/resize,w_1440,h_1440/quality,q_95/format,jpg" alt="" original-src="https://imageproxy.pbkrs.com/https://bang.xuangubao.cn/MTYxNDk1MzEzOQ=="/>
Repository address: https://github.com/aw31/openai-imo-2025-proofs/blob/main/README.md
Now let's take a look at the specific solution process for the first five problems with this new model.
The first problem is a coordinate geometry problem.
<img src="https://imageproxy.pbkrs.com/https://bang.xuangubao.cn/LTE2NzE2OTIwMjA=?x-oss-process=image/auto-orient,1/interlace,1/resize,w_1440,h_1440/quality,q_95/format,jpg" alt="" original-src="https://imageproxy.pbkrs.com/https://bang.xuangubao.cn/LTE2NzE2OTIwMjA="/>
It can be seen that the key to solving this problem is to find the n lines covering the points and the possible number of sunlight lines.
<img src="https://imageproxy.pbkrs.com/https://bang.xuangubao.cn/MTIyNTU1MTY3Mg==?x-oss-process=image/auto-orient,1/interlace,1/resize,w_1440,h_1440/quality,q_95/format,jpg" alt="" original-src="https://imageproxy.pbkrs.com/https://bang.xuangubao.cn/MTIyNTU1MTY3Mg=="/>
The model adopts a unique approach to determine all non-negative integers k that meet the conditions.
Lemma: For n≥4, any covering of P_n with n lines must use one side of the triangle.
A precise analysis of the case for n=3.
For the general case of n≥3, it is proven that for each n, there exists a configuration with k=0, 1, or 3.
The main conclusion is proven using the reduction lemma.
<img src="https://imageproxy.pbkrs.com/https://mmbiz-qpic.wscn.net/sz_mmbiz_png/UicQ7HgWiaUb20cxia8UPIzqz5ldrrd6AtfDSd9Wp5DlmkZ5iaia0tsLJuMkZx46OfyeN8uZtnWnttv8UnlCRfZiajyA/640/query-ZnJvbT1hcHBtc2cmd3hfZm10PXBuZw?x-oss-process=image/auto-orient,1/interlace,1/resize,w_1440,h_1440/quality,q_95/format,jpg" alt="" original-src="https://imageproxy.pbkrs.com/https://mmbiz-qpic.wscn.net/sz_mmbiz_png/UicQ7HgWiaUb20cxia8UPIzqz5ldrrd6AtfDSd9Wp5DlmkZ5iaia0tsLJuMkZx46OfyeN8uZtnWnttv8UnlCRfZiajyA/640/query-ZnJvbT1hcHBtc2cmd3hfZm10PXBuZw"/><img src="https://imageproxy.pbkrs.com/https://bang.xuangubao.cn/MTE4ODg5NTAwNA==?x-oss-process=image/auto-orient,1/interlace,1/resize,w_1440,h_1440/quality,q_95/format,jpg" alt="" original-src="https://imageproxy.pbkrs.com/https://bang.xuangubao.cn/MTE4ODg5NTAwNA=="/>
The schematic diagram is as follows.
<img src="https://imageproxy.pbkrs.com/https://bang.xuangubao.cn/NjYzOTgzMTQ=?x-oss-process=image/auto-orient,1/interlace,1/resize,w_1440,h_1440/quality,q_95/format,jpg" alt="" original-src="https://imageproxy.pbkrs.com/https://bang.xuangubao.cn/NjYzOTgzMTQ="/>
In the solution provided by the expert netizen, it is quite difficult to directly prove that the line passing through point H and parallel to AP is tangent to the circumcircle of triangle BER.
However, this problem can be rephrased: define X as the midpoint of EF on the side that does not contain B. If it can be proven that HX is parallel to EF, this conclusion can basically be established.
On the other hand, if this line is tangent, then it must touch the circle at the midpoint.
Therefore, it is sufficient to prove these two points.
<img src="https://imageproxy.pbkrs.com/https://bang.xuangubao.cn/MjU0Mzc4OTM2?x-oss-process=image/auto-orient,1/interlace,1/resize,w_1440,h_1440/quality,q_95/format,jpg" alt="" original-src="https://imageproxy.pbkrs.com/https://bang.xuangubao.cn/MjU0Mzc4OTM2"/>
During the proof process, the model mainly completes the following four steps:
<ol>
<li>Analytical settings and parameters. Let P be the circumcenter of triangle ACD. Connect AP and the intersection points of the two circles Ω and Γ are points E and F, respectively. Find the circumcircle equation passing through points B, E, and F.</li>
</ol>
<img src="https://imageproxy.pbkrs.com/https://mmbiz-qpic.wscn.net/sz_mmbiz_png/UicQ7HgWiaUb20cxia8UPIzqz5ldrrd6AtfOS9aEJNCqxDEuMv3hfQVhlt1FhMY37uAGPA6MR4D3CbfaNY3ItLBog/640/query-ZnJvbT1hcHBtc2cmd3hfZm10PXBuZw?x-oss-process=image/auto-orient,1/interlace,1/resize,w_1440,h_1440/quality,q_95/format,jpg" alt="" original-src="https://imageproxy.pbkrs.com/https://mmbiz-qpic.wscn.net/sz_mmbiz_png/UicQ7HgWiaUb20cxia8UPIzqz5ldrrd6AtfOS9aEJNCqxDEuMv3hfQVhlt1FhMY37uAGPA6MR4D3CbfaNY3ItLBog/640/query-ZnJvbT1hcHBtc2cmd3hfZm10PXBuZw"/>
Slide up and down to view
<img src="https://imageproxy.pbkrs.com/https://bang.xuangubao.cn/MTQ3ODI1Mzg4MA==?x-oss-process=image/auto-orient,1/interlace,1/resize,w_1440,h_1440/quality,q_95/format,jpg" alt="" original-src="https://imageproxy.pbkrs.com/https://bang.xuangubao.cn/MTQ3ODI1Mzg4MA=="/>
In the solution of the human player, the key to this problem is to prove the yellow inequality in the diagram below.
<img src="https://imageproxy.pbkrs.com/https://bang.xuangubao.cn/OTM3OTY3ODk3?x-oss-process=image/auto-orient,1/interlace,1/resize,w_1440,h_1440/quality,q_95/format,jpg" alt="" original-src="https://imageproxy.pbkrs.com/https://bang.xuangubao.cn/OTM3OTY3ODk3"/><img src="https://imageproxy.pbkrs.com/https://bang.xuangubao.cn/LTExNTEyNDkwNDk=?x-oss-process=image/auto-orient,1/interlace,1/resize,w_1440,h_1440/quality,q_95/format,jpg" alt="" original-src="https://imageproxy.pbkrs.com/https://bang.xuangubao.cn/LTExNTEyNDkwNDk="/>
The solution to this problem by the model is divided into the following four steps.
Consider bonza-type functions, that is, functions that satisfy property P_f.
When the function takes a value greater than 1 at a certain prime number, consider the congruence relation in the sense of modulo prime.
Conclusion: If a certain odd prime p satisfies f &gt; 1, then the entire function f must be an identity function.
Enter the main structural lemma: analyze the structure of the function in the case of non-identity functions.
<img src="https://imageproxy.pbkrs.com/https://mmbiz-qpic.wscn.net/sz_mmbiz_png/UicQ7HgWiaUb20cxia8UPIzqz5ldrrd6Atfp0LjqRAZENIVGO5Xrze6PNnGiaUfNxcicMrS4nw1cibFrJ069qlbkd7Zg/640/query-ZnJvbT1hcHBtc2cmd3hfZm10PXBuZw?x-oss-process=image/auto-orient,1/interlace,1/resize,w_1440,h_1440/quality,q_95/format,jpg" alt="" original-src="https://imageproxy.pbkrs.com/https://mmbiz-qpic.wscn.net/sz_mmbiz_png/UicQ7HgWiaUb20cxia8UPIzqz5ldrrd6Atfp0LjqRAZENIVGO5Xrze6PNnGiaUfNxcicMrS4nw1cibFrJ069qlbkd7Zg/640/query-ZnJvbT1hcHBtc2cmd3hfZm10PXBuZw"/>
Scroll up and down to view
<img src="https://imageproxy.pbkrs.com/https://bang.xuangubao.cn/LTk2MDU2NTQyNw==?x-oss-process=image/auto-orient,1/interlace,1/resize,w_1440,h_1440/quality,q_95/format,jpg" alt="" original-src="https://imageproxy.pbkrs.com/https://bang.xuangubao.cn/LTk2MDU2NTQyNw=="/>
The key steps to solving it can be expressed in the following form.
<img src="https://imageproxy.pbkrs.com/https://bang.xuangubao.cn/LTE1NzE1NjUxMTc=?x-oss-process=image/auto-orient,1/interlace,1/resize,w_1440,h_1440/quality,q_95/format,jpg" alt="" original-src="https://imageproxy.pbkrs.com/https://bang.xuangubao.cn/LTE1NzE1NjUxMTc="/><img src="https://imageproxy.pbkrs.com/https://bang.xuangubao.cn/MTI4MzAyMTQwOQ==?x-oss-process=image/auto-orient,1/interlace,1/resize,w_1440,h_1440/quality,q_95/format,jpg" alt="" original-src="https://imageproxy.pbkrs.com/https://bang.xuangubao.cn/MTI4MzAyMTQwOQ=="/>
Scroll up and down to view
<img src="https://imageproxy.pbkrs.com/https://bang.xuangubao.cn/MTI2ODAzNjI0OA==?x-oss-process=image/auto-orient,1/interlace,1/resize,w_1440,h_1440/quality,q_95/format,jpg" alt="" original-src="https://imageproxy.pbkrs.com/https://bang.xuangubao.cn/MTI2ODAzNjI0OA=="/>
The solution to this problem includes the following three parts.
<img src="https://imageproxy.pbkrs.com/https://bang.xuangubao.cn/MTMyNDUxMDg2Mw==?x-oss-process=image/auto-orient,1/interlace,1/resize,w_1440,h_1440/quality,q_95/format,jpg" alt="" original-src="https://imageproxy.pbkrs.com/https://bang.xuangubao.cn/MTMyNDUxMDg2Mw=="/><img src="https://imageproxy.pbkrs.com/https://bang.xuangubao.cn/Njg5Mjg1NzU1?x-oss-process=image/auto-orient,1/interlace,1/resize,w_1440,h_1440/quality,q_95/format,jpg" alt="" original-src="https://imageproxy.pbkrs.com/https://bang.xuangubao.cn/Njg5Mjg1NzU1"/><img src="https://imageproxy.pbkrs.com/https://bang.xuangubao.cn/NTcwNDA3OTE5?x-oss-process=image/auto-orient,1/interlace,1/resize,w_1440,h_1440/quality,q_95/format,jpg" alt="" original-src="https://imageproxy.pbkrs.com/https://bang.xuangubao.cn/NTcwNDA3OTE5"/>
During the problem-solving process, the model discussed the following three situations.
When λ ≥ c (no defense), Alice always wins.
When λ &gt; c, Alice wins.
When λ &lt; c (c = 1/√2), Bazza wins.
<img src="https://imageproxy.pbkrs.com/https://bang.xuangubao.cn/NTk1OTk1NzQ0?x-oss-process=image/auto-orient,1/interlace,1/resize,w_1440,h_1440/quality,q_95/format,jpg" alt="" original-src="https://imageproxy.pbkrs.com/https://bang.xuangubao.cn/NTk1OTk1NzQ0"/>
Scroll up and down to view
The sixth question involves combinatorial mathematics of graphs.
<img src="https://imageproxy.pbkrs.com/https://bang.xuangubao.cn/LTEyOTMzMzkyMjc=?x-oss-process=image/auto-orient,1/interlace,1/resize,w_1440,h_1440/quality,q_95/format,jpg" alt="" original-src="https://imageproxy.pbkrs.com/https://bang.xuangubao.cn/LTEyOTMzMzkyMjc="/>
This recognized difficult problem has only 6 human participants who can solve it, and all six large models, including o3, o4-mini, Gemini 2.5 Pro, Grok-4, and DeepSeek-R1, scored zero, even this super strong model from OpenAI also failed
Currently, there is no large model in the world that can solve the sixth question.
An Easter Egg
When Alexander Wei announced this news, he used the image of &#34;Strawberry.&#34;
&#34;Strawberry&#34; was the code name used by OpenAI during its internal research to promote a brand new reasoning model project, which is the &#34;o&#34; series model we are now familiar with.
<img src="https://imageproxy.pbkrs.com/https://mmbiz-qpic.wscn.net/sz_mmbiz_png/UicQ7HgWiaUb20cxia8UPIzqz5ldrrd6AtfZZia8ibO156ZqjrMBoZSxTWKERgRHTnCMRh6BL52qbGqQnEJes9jqo5Q/640/query-ZnJvbT1hcHBtc2cmd3hfZm10PXBuZw?x-oss-process=image/auto-orient,1/interlace,1/resize,w_1440,h_1440/quality,q_95/format,jpg" alt="" original-src="https://imageproxy.pbkrs.com/https://mmbiz-qpic.wscn.net/sz_mmbiz_png/UicQ7HgWiaUb20cxia8UPIzqz5ldrrd6AtfZZia8ibO156ZqjrMBoZSxTWKERgRHTnCMRh6BL52qbGqQnEJes9jqo5Q/640/query-ZnJvbT1hcHBtc2cmd3hfZm10PXBuZw"/>
Author Introduction
<img src="https://imageproxy.pbkrs.com/https://bang.xuangubao.cn/MjAzMjU1NjUzNg==?x-oss-process=image/auto-orient,1/interlace,1/resize,w_1440,h_1440/quality,q_95/format,jpg" alt="" original-src="https://imageproxy.pbkrs.com/https://bang.xuangubao.cn/MjAzMjU1NjUzNg=="/>
Alex Wei is a research scientist at OpenAI, primarily focusing on large language models and reasoning. He has previously researched the intersection of machine learning, game theory, and algorithms.
He obtained his Ph.D. in Computer Science from the University of California, Berkeley, under the supervision of Nika Haghtalab, Michael I. Jordan, and Jacob Steinhardt; and received his bachelor's and master's degrees from Harvard University, under the guidance of Jelani Nelson and Scott Kominers.
He was a member of the FAIR team and contributed to building the first artificial intelligence to reach human-level performance in the game of Diplomacy—CICERO. This achievement was published in the journal Science in 2022.
Reference:
https://x.com/alexwei_/status/1946477742855532918
Risk Warning and Disclaimer
The market has risks, and investment requires caution. This article does not constitute personal investment advice and does not take into account the specific investment goals, financial situation, or needs of individual users. Users should consider whether any opinions, views, or conclusions in this article are suitable for their specific circumstances. Investment based on this is at their own risk

OpenAI

OpenAI's brand new general reasoning model won the gold medal at the 2025 IMO competition, solving 5 problems and scoring 35 points, far surpassing other models. This achievement is considered a significant breakthrough in AI reasoning technology, potentially disrupting traditional thinking chains. The model took the exam under the same conditions as humans and ultimately won with a high score, demonstrating its potential in complex tasks

OpenAI's mysterious new model wins the IMO 2025 gold medal! Conquering the peak of Olympiad mathematics, Silicon Valley is buzzing