Model record

Nemotron 3 Ultra 550B-A55B

reasoningOpen weights

NVIDIA · released Jun 4, 2026

Overall standing

Overall
46.5rank 38 of 63
46.5 is at approximately the 40th percentile of the ranked Overall field. The field spans 19.0 to 62.2, with a median of 49.4.
8 of 8 measured
Category standingsReasoning 65.4Coding 46.9Math 3.1Agentic 13.8
Reasoning
65.4rank 29 of 63
65.4 is at approximately the 55th percentile of the ranked Reasoning field. The field spans 31.6 to 73.3, with a median of 65.0.
4 of 4 measured
Coding
46.9rank 46 of 63
46.9 is at approximately the 28th percentile of the ranked Coding field. The field spans 10.4 to 72.4, with a median of 54.5.
2 of 2 measured
Math
3.1rank 37 of 63
3.1 is at approximately the 40th percentile of the ranked Math field. The field spans 0.0 to 32.3, with a median of 5.7.
1 of 1 measured
Agentic
13.8rank 46 of 63
13.8 is at approximately the 28th percentile of the ranked Agentic field. The field spans 3.3 to 33.4, with a median of 19.0.
1 of 1 measured
Model factsOpen weights · Price not listed

Model facts come from the provider page. Listed API prices carry a separate first-party source and check date.

Provider
NVIDIA
Released
Weights
Open weights
Context window
1,000,000 tokens
Input price / Mtok
Not listed
Output price / Mtok
Not listed
Price source
Not listed
Reasoning effort
reasoning

Scores and sources

Measured scores retain their source and retrieval date. Missing scores are not treated as zero.

Benchmark scores, evaluation settings, sources, and retrieval dates for Nemotron 3 Ultra 550B-A55B
BenchmarkCategoryScoreEvaluation settingsSource
Terminal-Bench v2.1Coding53.9Artificial Analysis independent run; 89 tasks; Terminus 2 on E2B; 3 repeats; pass@1; 250-episode cap; 2-hour timeoutOpen source (opens in a new tab)
τ³-BankingAgentic13.8Artificial Analysis independent run; 97 tasks; 5 repeats; backend-state pass@1; BM25 plus grep retrieval; 200-step capOpen source (opens in a new tab)
AA-LCRReasoning67.0Artificial Analysis independent run; 100 roughly 100k-token multi-document questions; 3 repeats; equality-checker pass@1; no toolsOpen source (opens in a new tab)
Humanity's Last ExamReasoning26.6Artificial Analysis independent run; 2,158 text-only questions; pass@1; GPT-4o (August 2024) equality checker; no toolsOpen source (opens in a new tab)
GPQA DiamondReasoning86.7Artificial Analysis independent run; 198 questions; 5 repeats; regex-graded pass@1; no toolsOpen source (opens in a new tab)
SciCodeCoding39.9Artificial Analysis independent run; 288 test subproblems; scientist background included; 3 repeats; subproblem pass@1Open source (opens in a new tab)
IFBenchReasoning81.4Artificial Analysis independent run; 294 single-turn prompts; 5 repeats; official loose evaluator; prompt-level pass@1Open source (opens in a new tab)
CritPtMath3.1Artificial Analysis independent run; 70 test challenges; 5 repeats; two-step answer parsing; official CritPt grader; pass@1; no toolsOpen source (opens in a new tab)