# AI gold-medal result at the 2026 IOI

> Source: https://aiwiki.ai/wiki/ai_ioi_gold_2026
> Updated: 2026-09-06
> Fact-checked: 2026-09-06
> Categories: AI Benchmarks, NVIDIA, Reasoning Models
> License: CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/) - attribute to "AI Wiki (aiwiki.ai)"
> Cite as: AI Wiki. "AI gold-medal result at the 2026 IOI." aiwiki.ai, 6 Sept 2026. https://aiwiki.ai/wiki/ai_ioi_gold_2026
> From AI Wiki (https://aiwiki.ai), the free encyclopedia of artificial intelligence. Reuse freely with attribution.

In September 2026, researchers at [NVIDIA](https://aiwiki.ai/wiki/nvidia) reported that a fine-tuned version of [Nemotron 3](https://aiwiki.ai/wiki/nemotron_3) Ultra had scored 535.4 out of 600 points on the problem set of the 2026 International Olympiad in Informatics (IOI), the annual programming competition for secondary-school students, which was held in Tashkent, Uzbekistan, from August 9 to 16, 2026. The gold-medal threshold at that contest was 361.12 points and the highest score by a human contestant was 498.27. The result was described in an arXiv paper posted on September 2, 2026, and in a post on NVIDIA AI's X account early on September 5. The paper states that the system "was not an official IOI contestant and the run was not supervised by IOI," that its score was not included in the official rankings, and that the evaluation is reported "as an unofficial, unsupervised benchmark." [1][2][3]

The paper, "Post-Training Language Models for Gold-Medal Performance in Coding Competitions," also describes the training pipeline behind two competitive-programming variants of Nemotron 3, called Nemotron-3-Nano-CC and Nemotron-3-Ultra-CC, and a test-time strategy called GenCorrect. On the 2025 IOI problem set, which the authors used as a development benchmark, the smaller Nano-CC model rose from 130 points to 468 points across the stages of the pipeline and the larger Ultra-CC reached 502, both above the 2025 gold threshold of 438.3. The authors say that, to their knowledge, the 2026 run is the first time an AI system has outscored the highest-scoring human contestant on an IOI problem set. [1]

## Background: the IOI and its 2026 edition

The International Olympiad in Informatics is one of the international science olympiads for secondary-school students. Each contestant works alone over two competition days, with three tasks and five hours per day; a task is worth up to 100 points, so the maximum score is 600. Tasks are split into subtasks with different input constraints, and a submission earns credit for every subtask whose test cases it passes, so partial solutions and slower algorithms can score. A contestant's final score for a subtask is the best score across all of their submissions. [4][5]

The 38th IOI took place in Tashkent, Uzbekistan, from August 9 to 16, 2026, the first time Uzbekistan had hosted the event. The opening ceremony was held on August 10 at the CAEx exhibition complex in Tashkent. According to the IOI's official statistics site, 375 contestants took part; a press release from Uzbekistan's Ministry of Digital Technologies, distributed after the event, put the number of "top young coders" at 386 from over 97 countries, with total attendance above 900 delegates and guests. The official contest site lists NVIDIA among the event's sponsors, alongside companies including Huawei, Hudson River Trading and Cognition. [4][6][7][8]

The contest rules published by the organizing committee specify the constraints that matter for comparing an AI system with the human field. Solutions must be written in C++ and may not exceed 50 KiB. A contestant may submit to each task at most once per minute (the limit is lifted in the last 15 minutes of a round) and at most 50 times in total. The grading system reports a score for each subtask after every submission, plus limited feedback on one test case per unsolved subtask (the first among its lowest-scoring test cases). Contestants are forbidden from accessing any machine on the network or the internet other than the contest system; the rules note that "even running a single 'ping' command is strictly prohibited." [5]

The medal thresholds and task statistics for IOI 2026, from the official statistics site, were as follows. [4][9]

| Award | Number awarded | Threshold (points) |
|---|---|---|
| Gold | 31 | 361.12 or more |
| Silver | 62 | 303.28 or more |
| Bronze | 94 | 228.80 or more |
| Honourable mention | 43 | not applicable |

| Day | Task | Average score | Full solutions |
|---|---|---|---|
| 1 | Ball Machine | 29.90 | 4 |
| 1 | Monuments | 19.41 | 2 |
| 1 | Tiling Game | 72.72 | 244 |
| 2 | Classroom Game | 36.39 | 0 |
| 2 | Magic City | 33.67 | 4 |
| 2 | Partition | 18.20 | 6 |

The top human score was 498.27 points (83.05 percent of the maximum), by Qiwen Xu of China; the second- and third-placed contestants scored 482.75 and 473.53. No contestant fully solved Classroom Game. [10]

For comparison, the 2025 IOI in Sucre, Bolivia (July 27 to August 3, 2025) had 330 contestants, a gold threshold of 438.30 points, and a top human score of 591.23; the 2024 IOI in Alexandria, Egypt (September 1 to 8, 2024) had 362 contestants and a gold threshold of 359.71. The 2025 problem set is the one most AI laboratories have used to report IOI results. Because medal thresholds are set from the human score distribution, they move with the difficulty of each year's tasks: the 2026 gold cutoff of 361.12 was much lower than 2025's 438.30 and close to 2024's 359.71. [11][12][22]

## Earlier AI results in competitive programming

Competitive programming became a standard test of [reasoning models](https://aiwiki.ai/wiki/reasoning_models) during 2024 and 2025, and the IOI in particular has been used by several laboratories as a reference point. The main reported results before September 2026 are summarized below; each is a laboratory's own claim, and the settings differ in ways that matter for comparison.

| Date | System | Contest | Reported result | Setting |
|---|---|---|---|---|
| Sep 2024 (reported Feb 2025) | OpenAI o1-ioi | IOI 2024 | 213 points, 49th percentile; 362.14 with 10,000 submissions | Live entry with IOI committee's permission; 10,000 candidates per problem, 50 submissions [13][22] |
| Feb 2025 | OpenAI o3 (early checkpoint) | IOI 2024 | 395.64 points, above the roughly 360-point gold threshold | Retrospective, 50-submission limit [13] |
| Aug 2025 | OpenAI ensemble of general-purpose reasoning models | IOI 2025 | Gold-medal-level score, sixth place among human contestants (per OpenAI's Sheryl Hsu, as reported by The Decoder) | Online track, same five-hour limit and 50-submission cap [14] |
| Sep 2025 | Google DeepMind, advanced Gemini 2.5 Deep Think | ICPC World Finals 2025 | 10 of 12 problems, gold-medal level | Remote online environment following ICPC rules, started 10 minutes after the human teams [15] |
| Sep 2025 | OpenAI general-purpose reasoning models | ICPC World Finals 2025 | 12 of 12 problems, "enough for a 1st-place human ranking" | Company statement on X [16] |
| Oct 2025 | GenCluster with gpt-oss-120b (NVIDIA) | IOI 2025 | Gold-medal score, described as the first with an open-weight model | Up to 5,000 generations per subtask, 50-submission limit [17] |
| Dec 2025 | DeepSeek-V3.2-Speciale | IOI 2025 | 492 of 600, gold; the paper says it "ranked 10th" | 500 candidates per problem, 50 submitted [18] |
| Mar 2026 | Nemotron-Cascade-2-30B-A3B (NVIDIA) | IOI 2025 | 439.28 of 600, gold; 507.66 "achievable within 5000 generations" | At most 2,000 generations per subtask [19] |
| Sep 2026 | Competition Nemotron-3-Ultra-CC (NVIDIA) | IOI 2026 | 535.4 of 600, above the 361.12 gold threshold and the 498.27 top human score | Live, prospective, unofficial and unsupervised single run [1] |

[OpenAI](https://aiwiki.ai/wiki/openai)'s February 2025 paper "Competitive Programming with Large Reasoning Models" described the first live IOI attempt. The company entered a specialized system, o1-ioi, in the 2024 contest in Alexandria, Egypt, with the IOI committee's agreement, generating 10,000 candidate solutions per problem and using hand-built clustering and selection to pick 50 submissions; it scored 213 points, in the 49th percentile. When the submission limit was relaxed to 10,000, the same system reached 362.14 points, just above that year's gold cutoff of roughly 360. A later, retrospective evaluation of an early [o3](https://aiwiki.ai/wiki/o3) checkpoint on the same problems scored 395.64 points under the 50-submission limit without contest-specific heuristics. The official gold threshold that year was 359.71. [13][22]

In August 2025, OpenAI said a system built from general-purpose reasoning models had reached a gold-medal-level score on the IOI 2025 problem set. According to The Decoder's account of statements by OpenAI's Sheryl Hsu, the system competed in an online track under the same five-hour limit and 50-submission cap as human contestants and ranked sixth overall, ahead of all but five of the 330 human participants. The Decoder's report is also the source that NVIDIA's later GenCluster paper cites for this result, which notes that "details about the models and selection procedures were not released publicly." [14][17]

At the 2025 ICPC World Finals in Baku on September 4, 2025, [Google DeepMind](https://aiwiki.ai/wiki/google_deepmind) reported that an advanced version of [Gemini 2.5 Deep Think](https://aiwiki.ai/wiki/gemini_2_5_deep_think) solved 10 of 12 problems in a remote online environment "following ICPC rules, under the guidance of the competition organizers," and OpenAI said its general-purpose reasoning models solved all 12. Those results are covered in [AI at the 2025 ICPC World Finals](https://aiwiki.ai/wiki/icpc_2025_ai). [15][16]

Open-weight models followed. In October 2025 an NVIDIA team including several of the authors of the 2026 paper published GenCluster, a test-time compute framework that reached an IOI 2025 gold-medal score with [gpt-oss](https://aiwiki.ai/wiki/gpt_oss)-120b by generating up to 5,000 candidate solutions per subtask and selecting submissions through behavioral clustering, tournament ranking and a round-robin submission strategy; the authors called it the first gold with an open-weight model. [17] In December 2025, [DeepSeek](https://aiwiki.ai/wiki/deepseek)'s V3.2 report stated that its high-compute DeepSeek-V3.2-Speciale variant scored 492 of 600 on IOI 2025 and "ranked 10th," sampling 500 candidates per problem and submitting the 50 with the longest reasoning traces. [18] In March 2026, NVIDIA's Nemotron-Cascade 2, a 30B mixture-of-experts model with 3B active parameters, reported 439.28 points on IOI 2025 within at most 2,000 generations per subtask, calling itself the second open-weight model after DeepSeek-V3.2-Speciale to reach gold level at the IMO, the IOI and the ICPC World Finals. [19]

A separate strand of work argued that headline medal claims overstate model ability. LiveCodeBench Pro, a benchmark of Codeforces, ICPC and IOI problems annotated by olympiad medalists and published in June 2025, found that the best model at the time scored 0 percent on its hard problems without external tools and that high scores were "largely driven by implementation precision and tool augmentation, not superior reasoning." [20]

## NVIDIA's competitive-programming pipeline

The September 2026 paper is by Aleksander Ficek and Sean Narenthiran (listed as equal contributors), Mehrzad Samadi, Somshubra Majumdar and Boris Ginsburg. It was posted to arXiv on September 2, 2026 under a CC BY 4.0 license. NVIDIA AI's X post credits "our researchers." [1][2][3]

The authors' stated motivation is that earlier gold-medal systems were "often closed, rely on specialized models, or combine changes in training data, post-training, model scale, and inference-time compute," so the contribution of each component was hard to isolate. Their pipeline has four parts: problem curation, [supervised fine-tuning](https://aiwiki.ai/wiki/supervised_fine-tuning) on synthetic reasoning traces, [reinforcement learning](https://aiwiki.ai/wiki/reinforcement_learning) with executable rewards, and an iterative test-time strategy. [1]

### Data curation and synthetic traces

The team curated 22,000 problems from 16 regional and international competition families spanning two decades, plus problems from online programming platforms. An automated pipeline packaged each problem into an executable evaluation environment with its statement, constraints, test cases, auxiliary files and reference solutions, and kept only environments that produced consistent verdicts on known correct and incorrect solutions. All IOI 2025, ICPC 2025 and LiveCodeBench Pro problems were excluded from training data and the corpus was deduplicated against them; the paper notes that IOI 2026 needed no retrospective contamination filtering because the system ran before the problems were public. [1]

Reasoning traces for supervised fine-tuning were generated with [DeepSeek V4-Flash](https://aiwiki.ai/wiki/deepseek_v4_flash): 1.2 million traces for the Nano model and 477,642 for the Ultra model, with more generations allocated to harder problems. The mixture also contains "self-improvement" traces in which the teacher is given a problem and an earlier solution and asked to improve it, which the authors say exposes the student models to the kind of refinement behavior used at inference time. [1]

### The two models

| Model | Base | Total / active parameters | Post-training | SFT data | Hardware for SFT |
|---|---|---|---|---|---|
| Nemotron-3-Nano-CC | Nemotron-3-Nano-30B-A3B | 30B / 3B | SFT (three epochs) then RL | 1,200,000 traces from DeepSeek-V4-Flash | 64 NVIDIA GB300 GPUs |
| Nemotron-3-Ultra-CC | Nemotron-3-Ultra-550B-A55B (RLVR-teacher checkpoint) | 550B / 55B | SFT only (one epoch) | 477,642 traces from DeepSeek-V4-Flash | 128 NVIDIA GB300 GPUs |

Both models were fine-tuned with a global batch size of 64 and sequence packing up to 262K tokens. The Ultra model was initialized from what the paper calls its "RLVR-teacher checkpoint," a checkpoint previously used to distill reasoning ability into the general-purpose Nemotron 3 Ultra. [1]

Reinforcement learning was applied only to the Nano model, "in part because of the substantial computational cost of running RL at Ultra's scale." After filtering for reliable and fast environments, the RL corpus contained 3,219 problems, split at the parent-problem level into 2,847 training and 372 validation problems. Training used NeMo RL with [GRPO](https://aiwiki.ai/wiki/grpo): each step sampled 16 rollouts for each of 64 prompts at temperature 1.0, generated C++17 solutions were compiled and executed, and a solution received a reward of 1 for full credit and 0 otherwise, with no partial rewards and no reference-policy KL penalty. Rollouts could run to 255,144 generated tokens. The checkpoint at step 39 was selected on the validation set. [1]

### GenCorrect

GenCorrect is the paper's test-time strategy. It runs for up to five rounds. In each round the model generates up to 200 candidate solutions in parallel, which are compiled locally; invalid outputs are filtered; a diversity-selection step chooses up to 10 cluster centers using a score-blind heuristic and a similarity measure over token shingles of the normalized programs, and assigns every candidate to its nearest center; the best-ranked candidate from each cluster is submitted; and the evaluator's per-subtask scores are accumulated and fed, together with three reference solutions chosen to preserve solved subtasks, target remaining gaps and maintain diversity, into the next round's prompt. Five rounds of 10 submissions match the IOI's limit of 50 submissions per problem. The authors contrast this with earlier systems that split problems into subtasks and attacked each separately; GenCorrect gives the model all subtasks and lets it decide which to work on, which they say needs "substantially fewer necessary generations per problem." For ICPC, where the judge's feedback is binary, rounds continue until the problem is solved or performance plateaus. [1]

## Results on IOI 2025, ICPC 2025 and LiveCodeBench Pro

The paper evaluates on IOI 2025, ICPC 2025 and LiveCodeBench Pro, with IOI results reported as raw scores out of 600 or as normalized percentages. Score@1 is single-sample performance and Score@200 is the best of 200 parallel samples; ICPC and LiveCodeBench Pro use pass@1. Final IOI and ICPC results are averaged over 1,000 runs. All comparison models were run through the authors' own harness rather than copied from the models' reports, so the baseline numbers below can differ from what those models' developers published. [1]

| Model | IOI 2025 Score@1 | ICPC 2025 pass@1 | LiveCodeBench Pro pass@1 |
|---|---|---|---|
| Nemotron-3-Nano-30B-A3B (base) | 21.7% | 16.9% | 17.6% |
| Nemotron-Cascade-2-30B-A3B | 37.2% | 42.0% | 45.6% |
| gpt-oss-120b | 40.7% | 45.8% | 66.4% |
| Qwen3.6-35B-A3B | 40.8% | 32.0% | 58.4% |
| Nemotron-3-Ultra-550B-A55B (base) | 45.5% | 54.0% | 72.6% |
| DeepSeek-V4-Flash | 55.3% | 65.8% | 69.5% |
| DeepSeek-V4-Pro | 56.8% | 69.6% | 78.2% |
| GLM-5.2 | 66.0% | 65.7% | 83.8% |
| Nemotron-3-Nano-CC | 48.5% | 51.0% | 71.6% |
| Nemotron-3-Ultra-CC | 50.7% | 57.4% | 74.5% |

Source: Table 1 of the paper; IOI Score@1 is a normalized percentage of 600. [1]

Supervised fine-tuning produced most of the gain for the Nano model. Over three epochs its IOI 2025 Score@1 rose from 21.7 percent to about 47 percent (the paper reports 47.3 percent in one section and 46.7 percent, or 280 points, in another), with most of the improvement in the first epoch; reinforcement learning then added a smaller increment to 48.5 percent (291 points). The authors attribute the modest RL gain to the strong SFT starting point and to the difficulty of the setting: with a binary reward, a problem only provides a learning signal when its rollout group contains both successful and failed solutions, and rollouts of up to 255K tokens receive a single terminal reward. RL started from the base model without SFT improved Score@1 from 21.7 to 24.9 percent after 30 steps but did not catch up with SFT-initialized runs. [1]

The Ultra model gained less from its single SFT epoch, moving from 45.5 to 50.7 percent on IOI 2025, but the SFT-only Ultra-CC still beat the fully post-trained Nano-CC on all three benchmarks after roughly 478,000 examples, versus three epochs over 1.2 million examples plus RL for Nano-CC. The authors' conclusion is that "adapting a stronger base model with limited SFT can outperform extensive post-training of a smaller model" when model size and inference cost are not the main constraints. [1]

The test-time strategy contributed the largest raw gains. On IOI 2025, Nano-CC's mean score rose from 360.6 after the first GenCorrect round to 468.2 after five rounds, and Ultra-CC's from 343.9 to 502.0; Ultra-CC crossed the 438.3 gold threshold after three rounds and Nano-CC after four. The paper notes that at Score@1 the two models are only 2.2 percentage points apart (about 13 raw points) but at Score@200 Ultra-CC reaches 505 versus 461 for Nano-CC, and the gap widens further under GenCorrect. On ICPC 2025, Nano-CC went from 8.6 to 9.4 problems solved on average and Ultra-CC from 9.0 to 9.6, with both plateauing after two or three rounds; nine solved problems matched the fourth-placed gold-medal team at the human contest. [1]

| Stage | Nano-CC on IOI 2025 (points) | Ultra-CC on IOI 2025 (points) |
|---|---|---|
| Base model, Score@1 | 130 | 45.5% (about 273) |
| After SFT, Score@1 | 280 | 304 (50.7%) |
| After RL, Score@1 | 291 | not applied |
| Score@200 | 461 | 505 |
| GenCorrect, round 1 | 360.6 | 343.9 |
| GenCorrect, round 5 | 468.2 | 502.0 |
| 2025 gold threshold | 438.3 | 438.3 |

The paper reports the base Ultra model's IOI 2025 Score@1 only as a percentage (45.5 percent); the raw figure in parentheses is that percentage of 600. All other figures are the paper's own. [1]

## The IOI 2026 live run

After the general experiments, the team built a competition-specific system and evaluated it "prospectively on the IOI 2026 problem set during the official competition and before the problems were publicly available." The contest comprised two five-hour sessions on two days with three problems each. The paper says the system operated "under the same time, internet-access, and submission constraints as human contestants": no internet access, local code execution permitted, and up to 50 submissions per problem at a rate of one per minute. The live deployment used a peak allocation of up to 760 NVIDIA GB300 GPUs. [1]

### Competition-specific adaptations

Three changes were made for the live run, each chosen using IOI 2025 as a development benchmark. [1]

First, the Ultra model was fine-tuned on traces from [GLM-5.2](https://aiwiki.ai/wiki/glm_5_2) rather than DeepSeek-V4-Flash. The team compared the two as teachers on IOI 2025 and found GLM-5.2 scored higher with shorter outputs, a property that carried over after fine-tuning; shorter outputs allow more candidates to be generated within a fixed inference window. The RL stage was again skipped because RL at Ultra scale "exceeds our available compute budget." [1]

| System | IOI 2025 Score@1 | Mean generation length (tokens) |
|---|---|---|
| GLM-5.2 (teacher) | 66.0% | 85,927 |
| DeepSeek-V4-Flash (teacher) | 55.3% | 120,456 |
| Ultra-CC trained on GLM-5.2 data | 59.4% | 84,244 |
| Ultra-CC trained on DeepSeek-V4-Flash data | 50.7% | 89,626 |

Second, the final GenCorrect round was expanded. The first four rounds kept the standard 200 generations and 10 submissions per problem, but the fifth round generated 1,000 candidates and used an execution-based selection procedure, adapted from GenCluster, to pick the last 10 submissions: the model was prompted to write 50 test-input generators and validators, these were run until 100 valid inputs were obtained, every compiled candidate was executed on those inputs, and a model-written scoring script based on the problem's subtask criteria ranked the candidates. The authors' reasoning was that Ultra-CC's scores had begun to saturate across standard rounds in the IOI 2025 and ICPC 2025 experiments, so a single live attempt was better served by a larger final pool. [1]

Third, the Ultra model was quantized to [NVFP4](https://aiwiki.ai/wiki/nvfp4) with the NVIDIA Model Optimizer recipe, calibrated on 1,000 sequences of 32,768 tokens from the SFT mixture, to raise inference throughput. The configuration chosen for the live run (NVFP4 weights, FP8 KV cache, prefix caching off, multi-token prediction of 5) scored 52.8 percent Score@1 on IOI 2025 at 736.8 tokens per second per GPU, against 59.4 percent at 199.1 tokens per second per GPU for the BF16 model: the paper describes this as sacrificing 6.6 percentage points of single-sample accuracy for a 3.7-fold throughput increase, "enabling the large candidate batches required by GenCorrect within the competition window." [1]

### Result

| System | Score (of 600) | Medal level |
|---|---|---|
| Top human contestant | 498.27 | Gold |
| Gold-medal threshold | 361.12 | Gold |
| Competition Ultra-CC, live run | 535.40 | Gold |
| Ultra-CC, general five-round GenCorrect pipeline (post-competition, mean of 5 runs) | 521.72 (range 495.0 to 545.8) | Gold |

Source: Table 4 of the paper and the IOI 2026 official results. [1][4][10]

The live run scored 535.4, which the paper describes as 174.3 points above the gold threshold and 37.1 points above the top human contestant, "obtained from a single prospective run." To put the live number in context, the team later ran the standard five-round GenCorrect pipeline (without the competition-specific final round) five times on the same problems; those runs averaged 521.72 points with a range of 495.0 to 545.8, so the live result was 13.68 points above the mean of the general pipeline and inside its observed range. The authors note that the general pipeline's mean also exceeds both the gold threshold and the top human score. Per-problem scores for the live run are not given in the paper's text. [1]

The paper's acknowledgments thank George Armstrong, Wei Du and Igor Gitman "for sharing their insights from IMO," and "the IOI organization for their time and support." [1]

## How NVIDIA described the result

NVIDIA AI's post on X, published at 00:28 UTC on September 5, 2026, said the model scored 535.4 out of 600 "as graded by the IOI team" and "higher than the top-scoring human participant," that "the team competed unofficially in Uzbekistan, where the International Technical Committee supervised the human contestants," and that the model "had no internet access and faced the same time limits and submission constraints, using the same platform in parallel with the official competition." It linked to the technical report. [3]

The paper's own framing is more restrictive on one point. Its footnote states: "Our system was not an official IOI contestant and the run was not supervised by IOI. Therefore, its score was not included in the official rankings and the evaluation is reported as an unofficial, unsupervised benchmark." The X post says the committee supervised the human contestants, which is consistent with the paper, but a reader could take the phrase "using the same platform in parallel with the official competition" and "as graded by the IOI team" to imply oversight of the AI run; the paper says there was none. The paper does support the claims that the run used the same submission platform, time limits and submission constraints and had no internet access. [1][3]

## What the result does and does not show

The sources support the following reading of the result.

The run was prospective and contamination-free by construction. The paper states the system was executed during the live contest, before the problems were released, so the IOI 2026 problems could not have been in its training data. This is a stronger guarantee than the retrospective evaluations on IOI 2024 and IOI 2025 that most earlier claims relied on, which depended on exclusion and deduplication steps. [1]

The constraints matched the human rules on time, submissions and internet access, but not on resources. The paper is explicit that the result "should therefore be interpreted as a system-level comparison under the same time and submission limits, rather than an equal-resource comparison with human contestants." The system used up to 760 GB300 GPUs at peak, generated up to 200 candidates per round and 1,000 in the final round for each problem, and executed candidates locally on model-generated tests, none of which is available to a student at a single workstation. [1]

The score is unofficial and unsupervised. The system was not a contestant, its score is not in the official results, and the IOI did not supervise the run; the paper reports it as a benchmark. None of the sources consulted for this article (the paper, the X post, and the official IOI 2026 site and statistics pages as read on September 5, 2026) describes a verification of the AI run by the contest's committees. The X post's "as graded by the IOI team" is NVIDIA's description; the paper does not use that phrase, and its closest statement is a thank-you to "the IOI organization for their time and support." [1][3][10]

It is a single run. The live score of 535.4 is one sample. The paper's own post-competition runs of the general pipeline on the same problems ranged from 495.0 to 545.8, which shows how much variance the method has; the mean of those runs would still have exceeded the top human score. [1]

It is a comparison to one problem set, not to the person. The paper's claim is that the system outscored "the highest-scoring human contestant on an IOI problem set." The human contestants did not have the benefit of a second chance, parallel sampling or thousands of compiled candidates, and the human medal thresholds were set from the human score distribution alone. [1][4]

The models are not general-purpose in the sense OpenAI used for its 2025 results. Nano-CC and Ultra-CC are fine-tuned specifically on competitive-programming data, and the competition system was further tuned on IOI 2025. The paper says its "findings may not generalize beyond competitive programming." OpenAI said its 2025 IOI and ICPC results came from general-purpose reasoning models without contest-specific training, though it did not publish a technical report for either. [1][14][16]

The training corpus is not released. The authors say they cannot release the full training corpus "because of third-party redistribution restrictions," but plan to release the competition Ultra-CC checkpoint with runnable inference and evaluation recipes through NeMo-Skills. As of September 5, 2026, the releases index in the NeMo-Skills documentation listed six earlier releases (OpenCodeReasoning, OpenMathInstruct-2, OpenMathReasoning, OpenReasoning, Nemotron-Math-v2 and Nemotron-Math-Proofs) and no competitive-programming checkpoint. [1][21]

## Timeline

| Date | Event |
|---|---|
| Sep 1 to 8, 2024 | IOI 2024 in Alexandria, Egypt; OpenAI's o1-ioi competes live, scoring 213 points (49th percentile) [13][22] |
| Feb 3, 2025 | OpenAI publishes "Competitive Programming with Large Reasoning Models," including o3's retrospective 395.64 on IOI 2024 [13] |
| Jun 13, 2025 | LiveCodeBench Pro paper argues frontier models still fail hard olympiad problems without tools [20] |
| Jul 27 to Aug 3, 2025 | IOI 2025 in Sucre, Bolivia; gold threshold 438.30 [11] |
| Aug 2025 | OpenAI reports a gold-medal-level score and a sixth-place equivalent at IOI 2025 [14] |
| Sep 4, 2025 | ICPC World Finals in Baku; DeepMind (10 of 12) and OpenAI (12 of 12) announce results on Sep 17 [15][16] |
| Oct 16, 2025 | GenCluster paper: IOI 2025 gold with gpt-oss-120b [17] |
| Dec 2, 2025 | DeepSeek-V3.2 report: Speciale scores 492 on IOI 2025 [18] |
| Mar 19, 2026 | Nemotron-Cascade 2 report: 439.28 on IOI 2025 with a 30B-A3B model [19] |
| Aug 9 to 16, 2026 | IOI 2026 in Tashkent; NVIDIA runs its competition Ultra-CC system during the contest days [1][4] |
| Sep 2, 2026 | arXiv 2609.02849 posted, reporting 535.4 of 600 [2] |
| Sep 5, 2026 (00:28 UTC) | NVIDIA AI announces the result on X [3] |

## See also

- [AI gold medals at the 2025 IMO](https://aiwiki.ai/wiki/ai_imo_gold_2025)
- [AI at the 2025 ICPC World Finals](https://aiwiki.ai/wiki/icpc_2025_ai)
- [Nemotron](https://aiwiki.ai/wiki/nemotron) and [Nemotron 3.5 Lightning](https://aiwiki.ai/wiki/nemotron_3_5_lightning)
- [LiveCodeBench](https://aiwiki.ai/wiki/livecodebench)
- [OlympiadBench](https://aiwiki.ai/wiki/olympiadbench)
- [Test-time compute](https://aiwiki.ai/wiki/test_time_compute)
- [DeepSeek V4-Pro](https://aiwiki.ai/wiki/deepseek_v4_pro)
- [Gemini](https://aiwiki.ai/wiki/gemini)

## Coverage

Third-party coverage was sparse in the first days after the paper appeared. Tech Times published an explainer on September 5, 2026 that repeated the paper's headline figures (535.4 out of 600, a 37-point margin over the top human score of 498.27, the 361.12 gold threshold), used the Uzbek ministry's figure of 386 contestants, and framed the open question as "whether the result holds under independent verification".[23]

## References

1. [Post-Training Language Models for Gold-Medal Performance in Coding Competitions (HTML, v1)](https://arxiv.org/html/2609.02849v1) - arXiv (Aleksander Ficek, Sean Narenthiran, Mehrzad Samadi, Somshubra Majumdar, Boris Ginsburg), September 2, 2026.
2. [Post-Training Language Models for Gold-Medal Performance in Coding Competitions (abstract page, arXiv:2609.02849)](https://arxiv.org/abs/2609.02849) - arXiv, September 2, 2026.
3. [Post: "Congrats to our researchers for exceeding the gold medal threshold on the International Olympiad in Informatics (IOI) 2026 problem set"](https://x.com/NVIDIAAI/status/2096032566310789528) - NVIDIA AI on X, September 5, 2026.
4. [IOI 2026](https://stats.ioinformatics.org/olympiads/2026) - International Olympiad in Informatics statistics site, accessed September 5, 2026.
5. [IOI 2026 Contest Rules](https://ioi2026.uz/contest-rules) - IOI 2026 Organizing Committee, accessed September 5, 2026.
6. [International Olympiad in Informatics 2026, Uzbekistan](https://ioi2026.uz/) - IOI 2026 official site, accessed September 5, 2026.
7. [Opening ceremony of the 38th international olympiad in informatics took place](https://ioi2026.uz/news) - IOI 2026 official site, August 10, 2026.
8. [Uzbekistan Hosts 38th International Olympiad in Informatics, Cementing Role as Central Asia's Leading IT Hub](https://finance.yahoo.com/technology/articles/uzbekistan-hosts-38th-international-olympiad-130500356.html) - ACCESS Newswire via Yahoo Finance (Ministry of Digital Technologies of the Republic of Uzbekistan), August 18, 2026.
9. [IOI 2026: Tasks](https://stats.ioinformatics.org/tasks/2026) - International Olympiad in Informatics statistics site, accessed September 5, 2026.
10. [IOI 2026: Results](https://stats.ioinformatics.org/results/2026) - International Olympiad in Informatics statistics site, accessed September 5, 2026.
11. [IOI 2025](https://stats.ioinformatics.org/olympiads/2025) - International Olympiad in Informatics statistics site, accessed September 5, 2026.
12. [IOI 2025: Results](https://stats.ioinformatics.org/results/2025) - International Olympiad in Informatics statistics site, accessed September 5, 2026.
13. [Competitive Programming with Large Reasoning Models (arXiv:2502.06807)](https://arxiv.org/abs/2502.06807) - arXiv (OpenAI: Ahmed El-Kishky, Alexander Wei, Andre Saraiva, Borys Minaiev, Daniel Selsam, David Dohan, Francis Song, Hunter Lightman, Ignasi Clavera, Jakub Pachocki, Jerry Tworek, Lorenz Kuhn, Lukasz Kaiser, Mark Chen, Max Schwarzer, Mostafa Rohaninejad, Nat McAleese, o3 contributors, Oleg Mürk, Rhythm Garg, Rui Shu, Szymon Sidor, Vineet Kosaraju, Wenda Zhou), February 3, 2025.
14. [OpenAI's AI system wins a gold medal-level score at the International Olympiad in Informatics 2025](https://the-decoder.com/openais-ai-system-wins-a-gold-medal-level-score-at-the-international-olympiad-in-informatics-2025/) - The Decoder (Maximilian Schreiner), August 12, 2025.
15. [Gemini achieves gold-medal level at the International Collegiate Programming Contest World Finals](https://deepmind.google/blog/gemini-achieves-gold-medal-level-at-the-international-collegiate-programming-contest-world-finals/) - Google DeepMind, September 17, 2025.
16. [Post: "Our general-purpose reasoning models solved all 12 problems at the 2025 International Collegiate Programming Contest (ICPC) World Finals"](https://x.com/OpenAI/status/1968368133024231902) - OpenAI on X, September 17, 2025.
17. [Scaling Test-Time Compute to Achieve IOI Gold Medal with Open-Weight Models (arXiv:2510.14232)](https://arxiv.org/abs/2510.14232) - arXiv (Mehrzad Samadi, Aleksander Ficek, Sean Narenthiran, Siddhartha Jain, Wasi Uddin Ahmad, Somshubra Majumdar, Vahid Noroozi, Boris Ginsburg), October 16, 2025.
18. [DeepSeek-V3.2: Pushing the Frontier of Open Large Language Models (arXiv:2512.02556)](https://arxiv.org/abs/2512.02556) - arXiv (DeepSeek-AI), December 2, 2025.
19. [Nemotron-Cascade 2: Post-Training LLMs with Cascade RL and Multi-Domain On-Policy Distillation (arXiv:2603.19220)](https://arxiv.org/abs/2603.19220) - arXiv (Zhuolin Yang, Zihan Liu, Yang Chen, Wenliang Dai, Boxin Wang, Sheng-Chieh Lin, Chankyu Lee, Yangyi Chen, Dongfu Jiang, Jiafan He, Renjie Pi, Grace Lam, Nayeon Lee, Alexander Bukharin, Mohammad Shoeybi, Bryan Catanzaro, Wei Ping), March 19, 2026.
20. [LiveCodeBench Pro: How Do Olympiad Medalists Judge LLMs in Competitive Programming? (arXiv:2506.11928)](https://arxiv.org/abs/2506.11928) - arXiv (Zihan Zheng, Zerui Cheng, Zeyu Shen, Shang Zhou, Kaiyuan Liu, Hansen He, Dongruixuan Li, Stanley Wei, Hangyi Hao, Jianzhu Yao, Peiyao Sheng, Zixuan Wang, Wenhao Chai, Aleksandra Korolova, Peter Henderson, Sanjeev Arora, Pramod Viswanath, Jingbo Shang, Saining Xie), June 13, 2025.
21. [NVIDIA-NeMo/Skills](https://github.com/NVIDIA-NeMo/Skills) - GitHub (NVIDIA), accessed September 5, 2026.
22. [IOI 2024](https://stats.ioinformatics.org/olympiads/2024) - International Olympiad in Informatics statistics site, accessed September 5, 2026.
23. [NVIDIA AI Outscored Every Human at IOI 2026: How GenCorrect Made It Possible](https://www.techtimes.com/articles/326744/20260905/nvidia-ai-outscored-every-human-ioi-2026-how-gencorrect-made-it-possible.htm) - Tech Times (Roger Satterfield), September 5, 2026.