Verification score
Share of 80 questions answered correctly · higher is better
Google Anthropic OpenAI SpaceXAI DeepSeek Meta Qwen MoonshotAI Other labs
ResearchLLM Verifications Leaderboard
Data as of · 78 models, 80 questions · reviewed by Suhas Bhairav · methodology
LLM leaderboard
Score vs. cost per question
Filled dots: no cheaper model scores as high · Claude Opus 5.5 has the top score
Google Anthropic OpenAI SpaceXAI DeepSeek Meta Qwen MoonshotAI Other labs
Output speed
Median output tokens per second · higher is better
Cost per question
Average billed USD, reasoning included · lower is better
Time to first token
Median seconds to the first token · lower is better
Results by category
Questions answered correctly in each category · categories hold 6 to 30 questions, so read single cells with care
| Model | Score | Science30 questions | Engineering15 questions | Medicine11 questions | Economics9 questions | Law6 questions | Humanities & social science9 questions |
|---|---|---|---|---|---|---|---|
| Claude Opus 5.5Anthropic | 85.0 | 28/30 | 12/15 | 7/11 | 8/9 | 5/6 | 8/9 |
| Gemini 4 ArgonGoogle | 83.8 | 25/30 | 13/15 | 7/11 | 8/9 | 6/6 | 8/9 |
| GPT-5.3-CodexOpenAI | 82.5 | 28/30 | 11/15 | 7/11 | 8/9 | 4/6 | 8/9 |
| Gemini 3.8 FlashGoogle | 82.5 | 26/30 | 12/15 | 6/11 | 8/9 | 6/6 | 8/9 |
| Claude Fable 5.1Anthropic | 81.2 | 25/30 | 13/15 | 7/11 | 7/9 | 5/6 | 8/9 |
| Grok 4.7SpaceXAI | 80.0 | 27/30 | 11/15 | 6/11 | 8/9 | 4/6 | 8/9 |
| Muse WonderMeta | 78.8 | 26/30 | 11/15 | 5/11 | 8/9 | 4/6 | 9/9 |
| Claude Sonnet 5.5Anthropic | 78.8 | 24/30 | 13/15 | 7/11 | 8/9 | 3/6 | 8/9 |
| Gemini 3.1 Pro PreviewGoogle | 77.5 | 26/30 | 10/15 | 7/11 | 8/9 | 3/6 | 8/9 |
| Muse Spark OpenMeta | 77.5 | 25/30 | 11/15 | 6/11 | 8/9 | 4/6 | 8/9 |
| Kimi K3.1 ProMoonshotAI | 77.5 | 26/30 | 11/15 | 6/11 | 8/9 | 3/6 | 8/9 |
| GPT-6 AstraOpenAI | 76.2 | 26/30 | 10/15 | 6/11 | 8/9 | 3/6 | 8/9 |
| GPT-6.1 Terra ProOpenAI | 76.2 | 26/30 | 10/15 | 6/11 | 8/9 | 3/6 | 8/9 |
| GPT-6.1 SolOpenAI | 76.2 | 27/30 | 10/15 | 5/11 | 8/9 | 3/6 | 8/9 |
| GPT-5.6 TerraOpenAI | 75.0 | 26/30 | 11/15 | 4/11 | 7/9 | 4/6 | 8/9 |
| Gemini 3.5 Flash LiteGoogle | 75.0 | 23/30 | 10/15 | 7/11 | 7/9 | 4/6 | 9/9 |
| Qwen3.8 Max (0902)Qwen | 75.0 | 24/30 | 9/15 | 6/11 | 8/9 | 4/6 | 9/9 |
| GPT-6.1 TerraOpenAI | 73.8 | 26/30 | 9/15 | 6/11 | 7/9 | 3/6 | 8/9 |
| DeepSeek V4.1 Flash 1017DeepSeek | 73.8 | 26/30 | 9/15 | 4/11 | 7/9 | 4/6 | 9/9 |
| GPT-6 SolOpenAI | 73.8 | 27/30 | 8/15 | 5/11 | 7/9 | 4/6 | 8/9 |
Which model for which job
Computed from this run by the rule in each label
- Highest verification scoreClaude Opus 5.585.0% · 68/80
- Fastest with a score of 80+Gemini 3.8 Flash142 tok/s · 82.5%
- Cheapest with a score of 80+Gemini 3.8 Flash$0.0037 per question · 82.5%
- Quickest to respond with a score of 80+GPT-5.3-Codex1.4 s to first token
- Best open-weights modelMuse Spark Open77.5%
Quick answers
- Which AI model is the most accurate?
- Claude Opus 5.5 by Anthropic scored highest in the 2026-10-27 run, answering 68 of 80 questions correctly (85.0%), out of 78 models tested.
- Which AI model is the fastest?
- Nemotron 3.5 Super VL by NVIDIA had the highest median output speed in the 2026-10-27 run: 715 tokens per second.
- Which AI model is the cheapest to run?
- Granite 4.0 Micro by IBM had the lowest billed cost per question in the 2026-10-27 run: $0.000051, reasoning tokens included.
- Which AI model responds the quickest?
- Ministral 3 3B 2512 by Mistral had the shortest median time to first token in the 2026-10-27 run: 0.3 s.
- How are these benchmarks measured?
- Every model is called through OpenRouter's public API with the same 80 questions, each with one correct answer, plus 3 streamed speed runs. Every request and response is stored and hashed.