Cognixa Labs
Line 1 research office · cognion.io

How we measure.

Cognion, the lab's benchmark platform, scores frontier models across intelligence, coding, and math. The methodology is public, the runs are logged, and the leaderboard is free to read. This page explains the system; the live data lives at cognion.io.

The method, in four rules

Published, reproducible, independent
1

Published harness.

The evaluation harness is public. Every prompt set, scoring script, and environment note ships with the results, so a score is an experiment, not an assertion.

2

Logged runs.

Each evaluation run is recorded with its configuration and outputs. If we cannot reproduce a number on request, it does not stay on the board.

3

Three dimensions, one composite.

Intelligence (MMLU), math (GSM8K), and coding (HumanEval) are reported separately and as a weighted composite. Single numbers simplify; we show what they hide.

Leaderboard snapshot

Intelligence Index via artificialanalysis.ai · live board at cognion.io
Frontier models: top of the boardRanked
ModelLabIntelligence IndexRank
Claude Opus 5.5Anthropic5801
Kimi K3.1 ProMoonshot AI5502
GPT-6 AstraOpenAI5303
Claude Fable 5.1Anthropic5304
Gemini 4 Argon (High)Google5305
GPT-6.1 SolOpenAI5206

Scores from the Artificial Analysis Intelligence Index, October 2026 snapshot. The live Cognion board at cognion.io carries every category, the full run log, and the current rankings.

What we publish

Free to read, always
Private trials

Need a model measured before you bet the quarter on it?

The lab runs private, pre-procurement evaluations on the same rails as the public board: your workloads, your data, our methodology, your report.

Trial termsConfidential
WorkloadsYours, not synthetic ones
ReportMethodology included
ConfidentialityResults stay yours
TurnaroundScoped per request