Live model pulse

When a model is in the run, how often is its answer the best?

Every consens.io run sends one question to several models. A judge model then picks the strongest single answer. Here you see, for each model family, how often that pick went to it, out of the runs it was actually in.

Best-answer rate

Share of a family's runs in which the judge picked its answer. Base and Pro models count as one family.

Period 7 days 30 days 90 days All time

Runs All Consensus Agent

Only runs with Sort by

01 OpenAI / ChatGPT 244 of 434 runs · 3.0× fair share 56%
02 DeepSeek 72 of 458 runs · 0.9× fair share 16%
03 Google / Gemini 70 of 508 runs · 0.7× fair share 14%
04 Meta / Muse 28 of 210 runs · 0.7× fair share 13%
05 Moonshot AI / Kimi 17 of 134 runs · 0.7× fair share 13%
06 Anthropic / Claude 31 of 283 runs · 0.6× fair share 11%
07 Z.ai / GLM 32 of 300 runs · 0.6× fair share 11%
08 xAI / Grok 29 of 378 runs · 0.4× fair share 8%
09 Mistral 3 of 142 runs · 0.1× fair share 2%

Too few runs to rank (under 10): Qwen 3, MiMo 1

526 judged runs since 2026-08-11 · 5.4 answers per run on average. No prompt text or user identity enters this tally.

How to read it

01

A rate, not a vote count

A family that is in more runs collects more picks without being any better. So each family's picks are divided by the runs it was in: 30 picks from 100 runs is a 30% best-answer rate.

02

The tick is chance

With six answers in a run, a blind draw would pick each one about 17% of the time. The tick on every bar marks that fair share for the runs the family was in. "2.8× fair share" means picked 2.8 times as often as chance.

03

Wide band, few runs

The shaded band is the range the true rate plausibly lies in. Families under 10 runs are listed but not ranked.

04

Pick a rival

"Only runs with" narrows the board to runs that included one family and shows each other family's head-to-head score against it.

05

Honest limits

The questions are real and uncontrolled, and the pick is a judgment, not a verified answer key. This is not a popularity vote and not an accuracy benchmark.

Two different lenses

Live behavior and controlled accuracy answer different questions.

Model pulse

Whose answer stands out in everyday runs?

Questions: Real, changing prompts
Signal: Judge picks per run a family was in
Best used for: Product behavior and model momentum

MMLU-Pro benchmark

What is correct under fixed conditions?

Questions: 314 known-answer exam items
Signal: Measured accuracy
Best used for: Controlled model and consensus comparison

Open the full benchmark

Run your own comparison

No model wins every time. So ask several.

Open app