OpenAI's reasoning system achieved a perfect 12/12 on the same problem at the 2025 ICPC World Finals, placing first according to official rules. DeepMind's Gemini 2.5 also achieved gold medal status. ICPC is a high-intensity algorithmic competition, and the results demonstrate that general reasoning models are approaching top human performance in complex search and engineering implementation. For detailed sources, see the references at the end of this article. I. Overview and Implications of the Event 1. Results and Competition System: The Value of a Full ICPC Score The ICPC World Finals lasted 300 minutes and consisted of 12 problems. Only fully correct answers were scored, and ranking was based on time. OpenAI's reasoning system achieved a perfect solution on the same problem, passing most of them on the first try. DeepMind achieved gold medal status on all 12 problems, further validating the integrated algorithmic and engineering capabilities of its large-scale model.
2. Pay attention to the boundaries: it is not "official victory on the spot"
This is an offline evaluation of the same question, and OpenAI and DeepMind are not included in the list as official participating teams. The real competition also includes dimensions such as team collaboration, fault recovery and stress management, and AI still needs systematic verification in these aspects.
(1) Key points of the competition
The total time is fixed, and the question types cover graph theory, number theory, geometry and data structure, with an extremely low error tolerance rate.
(2) Model performance details
OpenAI hit the most questions on the first try, and the most difficult questions were passed after multiple submissions; DeepMind demonstrated unique strategies for some difficult questions.
(3) Industry significance
From code agency to scientific research engineering, competition-level reasoning and search can be transferred to high-value scenarios such as defect location, constraint solving and automated verification.
II. Turning “competition-level reasoning” into productivity
1. Evaluation method: Business set alignment ICPC rules
Construct an enterprise evaluation set covering time limit, memory and provability, adopt a strong constraint and penalty strategy of “only giving full marks” to measure the stability and fallback path of the model on real difficult problems.
2. Engineering closed loop: Agent + tool chain + sandbox execution
Introduce problem decomposition templates, differential single testing and minimal edit repair, combined with restricted sandbox and auditable logs to ensure reproducibility and traceability.
(1) Problem decomposition and planning
Standardize problem meaning analysis, sample construction and boundary enumeration.
(2) Code generation and self-testing
Integrated compilation, sample regression and failure retry; introduce multi-solution voting to improve robustness.
（3） Resources and Security
Limit time, memory, and system calls to avoid unauthorized access and resource exhaustion.
a. Cost Control
Cache common subtasks and search results to reduce repeated inference overhead.
b. Reliability Indicators
Use pass rate, penalty time, and number of retries as core health scores.
c. Grayscale and Rollback
Preset model switches and quota alerts to reduce unpredictable fluctuations.
Frequently Asked Questions (Q&A)
Q: Did OpenAI “officially win”?
A: No. This is an offline evaluation of the same ICPC problem, not an official on-site ranking; however, a score of 12/12 is highly valuable under ICPC rules.
Q: How does DeepMind's Gemini 2.5 compare to OpenAI's reasoning system?
A: Gemini 2.5 reaches gold medal level and excels on individual problems, but its overall number of solved problems is lower than the OpenAI reasoning system's perfect score, demonstrating strong reasoning and engineering execution.
Q: What lessons can enterprises learn from the ICPC's challenges?
A: Strict time constraints and zero-tolerance scoring force systems to possess robust planning, rapid verification, and automated error correction capabilities, precisely addressing the reliability and auditability requirements of production environments.
Q: How can we quickly verify whether a model is worth migrating?
A: We can first construct an "ICPC-ized" evaluation set using a small sample of business applications to observe factual consistency, latency, and manual rework rate. If it consistently outperforms the existing baseline, we can then expand the phased coverage.
