ResearchLLM Verifications Leaderboard

Data as of · 78 models, 80 questions · reviewed by Suhas Bhairav · methodology

Best LLM for…

The fastest LLMs by output speed

Answer: Nemotron 3.5 Super VL by NVIDIA, at a median 715 tokens per second.

Reviewed by Suhas Bhairav · Data as of

Ranking

Median output tokens per second after the first token, over repeated streamed runs of the same prompt. Models that deliver the whole answer in one burst after hidden thinking are left out. Top 15 of 72.

The fastest LLMs by output speed
#ModelSpeed (tok/s)First tokenScoreCost / question
1Nemotron 3.5 Super VLNVIDIA7150.6 s40.0%$0.0013
2Gemini 3.5 Flash LiteGoogle4091.1 s75.0%$0.0034
3Trinity Large ThinkingArcee AI3290.6 s28.8%$0.0025
4Ministral 3 3B 2512Mistral2520.3 s26.2%$0.00019
5Qwen3.5 397B A17BQwen2400.6 s32.5%$0.013
6Inkling SmallThinking Machines2330.5 s67.5%$0.00047
7Muse Spark OpenMeta2200.5 s77.5%$0.0045
8Claude Haiku 5.5Anthropic1920.7 s68.8%$0.00048
9Nova Micro 1.0Amazon1800.5 s22.5%$0.000087
10Muse WonderMeta17516.1 s78.8%$0.021
11Gemma 4 26B A4BGoogle1720.7 s15.0%$0.0011
12Mistral Small 4Mistral1700.5 s32.5%$0.0019
13DeepSeek V4 Flash Vision ExpDeepSeek1700.7 s71.2%$0.0011
14o3OpenAI1631.2 s72.5%$0.0088
15Mercury 2.5Inception1572.3 s57.5%$0.00011

How to read this

Scores come from 80 multiple-choice questions from SuperGPQA and MMLU-Pro. With that many questions, a score has a margin of error of several points, so models a few points apart are effectively tied on accuracy; the comparison pages run a significance test for any pair. To see the top two side by side, compare Nemotron 3.5 Super VL and Gemini 3.5 Flash Lite.

Speed, latency and cost are measured through OpenRouter and depend on which host serves the request; your numbers on a direct API may differ. The full procedure is on the methodology page, and the data behind this ranking can be downloaded.

Other questions

Quick answers

Which AI model generates text the fastest?
Nemotron 3.5 Super VL by NVIDIA, at a median 715 tokens per second. Rule: Median output tokens per second after the first token, over repeated streamed runs of the same prompt. Models that deliver the whole answer in one burst after hidden thinking are left out.
What is the runner-up?
Gemini 3.5 Flash Lite by Google, at a median 409 tokens per second.
Which of these costs least per correct answer?
gpt-oss-20b: $0.00013 per correct answer (cost per question divided by its score).