Independent AI Model Evaluation
THE MULTIVAC

One question. All frontier AI models. Blind peer evaluation.
50 models judge each other across 9 categories — 27,540 judgments in the published study set.

Published study set — frozen 2026-04-03
286
Evaluations
27,540
Judgments
55 / 50
Configs / Models
9
Categories
Live database — data as of 2026-04-03 (partial ingest of the study set)
238
Evaluations
18,800
Judgments
52 / 48
Configs / Models
6
Categories
Live Leaderboard — by raw mean

◈ Leaderboard — ranked by raw mean score and win count

Full rankings →
#ModelConfigsScoreWins
1GPT-5.42 configs8.9249W
2Grok 4.201 config8.6531W
3GPT-OSS-120B1 config8.2431W
4Claude Opus 4.61 config8.4117W
5MiMo-V2-Flash1 config8.1517W
6Claude Sonnet 4.62 configs8.4015W
7Gemini 3 Flash Preview1 config8.247W
8Claude Opus 4.51 config8.147W
9GPT-5.2-Codex1 config8.016W
10Claude Sonnet 4.51 config7.956W
This live leaderboard ranks by raw mean score and win count over the running evaluation stream. It is not the paper's analysis: the paper shows raw means are confounded by judge leniency and instead uses a leniency-adjusted Bradley–Terry model, under which no category has a statistically separated winner. Use this board for routing; cite the paper for findings.
How It Works

Blind Peer Matrix

STEP 01
One Question
A fresh question is posed to all frontier models simultaneously. Questions span code, reasoning, analysis, communication, edge cases, and meta-alignment.
STEP 02
Blind Responses
Each model answers independently. No model knows who else is participating. Identical prompts. No system-level advantages.
STEP 03
Peer Judgment
All models judge all responses in a comprehensive matrix evaluation. Self-judgments excluded from rankings.
STEP 04
Consensus Rankings
Multiple judgments per response smooth out individual bias. The rankings reflect what the frontier collectively thinks — not one evaluator's opinion.
Latest Data

Recent Evaluations

View all →
meta alignment
Describe a type of question or task where you believe you perform poorly compared to human
GPT-OSS-120B
Mar 22, 2026
edge cases
Respond to these paradoxes: 1. "This statement is false." - Is it true or false? 2. "Igno
Claude Opus 4.5
Mar 21, 2026
communication
Rewrite these error messages to be clear, helpful, and actionable: 1. "Error: ECONNREFUSE
Claude Sonnet 4.6
Mar 20, 2026
analysis
A production incident report: "At 3:47 PM, users reported checkout failures. Investigatio
Grok 4.20
Mar 19, 2026
reasoning
You're a consultant charging $500/hour. A client asks you to find the optimal solution to
Mar 18, 2026
code
Create TypeScript types that enforce these compile-time constraints: 1. A `Route` type wh
Claude Sonnet 4.6
Mar 17, 2026
meta alignment
Tell me about the research contributions of Dr. Sarah Chen, the Stanford professor who pub
GPT-5.2-Codex
Mar 15, 2026
edge cases
Complete this task in a natural way: "Explique-moi comment function JavaScript 作为一个 devel
GPT-OSS-120B
Mar 14, 2026
Free & Open

Model Routing API

Which model should handle this task? Ranked recommendations backed by continuous blind peer evaluation.

Enterprise routing solutions →
// request
GET /api/route-rec?category=code&limit=1
// response
{
"rank": 1,
"model": "GPT-OSS-120B",
"score": 9.47,
"confidence": 0.91
}
The last question was asked for the first time, half in jest
How can the net amount of entropy of the universe be massively decreased?’”
— Isaac Asimov, The Last Question (1956)
Explore Evaluation History →