Nemotron 3 Ultra has unusually substantial official documentation: an open 550B/55B-active MoE checkpoint, text-only I/O, 1M-context claim, configurable reasoning, tool use, and NIM deployment/API materials. Its technical score is below GPT-5.5 and Claude Opus 4.8 because independent frontier-comparative evidence is limited, but above GLM-4.6 on architecture, context, and deployment scope. LiveCodeBench reporting lists Nemotron3U at 89% Pass@1; NVIDIA also reports SWE-Bench Verified in its technical report, though the supplied evidence does not provide a directly comparable score here. Coding and reasoning are therefore scored below DeepSeek-V3.2 and Kimi K2.5 despite the favorable LiveCodeBench result. In particular, Agent Arena places it #42 of 44, a material contrary signal for end-to-end agent workflows. Text-only I/O keeps multimodal capability aligned with DeepSeek-V3.2 and well below multimodal anchors. Artificial Analysis reports 203.8 output tokens/s and $0.675/$2.675 per million input/output tokens, supporting stronger speed and cost scores than premium closed models, subject to provider and configuration variation. NVIDIA's NIM documentation, compatible APIs, checkpoints, and NeMo repository make developer experience comparatively strong. Pricing clarity is weaker because cited AI Enterprise pricing is platform-level rather than a Nemotron 3 Ultra-specific production tariff. Public adoption remains modest versus OpenAI and Anthropic anchors. Evidence quality is reasonable but not top-tier: official claims are well documented, while no DeepSWE result was found, no Terminal-Bench/Aider result is supplied, and the preference-family source is Agent Arena rather than LMArena or Arena-Hard.
