01 · Ask
Six models, independently
The same question goes to all six consensus models in parallel. Each answers on its own, with tools and web access switched off.
Benchmark · MMLU-Pro
We put the consensus to the test.
consens.io sends one question to six leading models and fuses their answers into a single consensus. So we asked the obvious question back: does that consensus actually beat the models it's built from? Here's the run, graded against known-correct answers, nothing hidden.
When the models disagree
Accuracy on the questions where they split · MMLU-Pro
72.1%
70.6%
69.1%
64.7%
58.8%
44.1%
39.7%
The blue bar is the consens.io consensus, first on the questions where the models can't agree.
How it ran
One question, six answers, one consensus.
MMLU-Pro is a hard, expert-level multiple-choice exam. Every question runs closed-book: no web search, no tools, no outside sources, just what each model knows.
02 · Fuse
The consensus is built
The exact same consensus logic the app uses combines the six answers into one final answer, with Gemini 3.5 Flash as the synthesizing model, the same path a real prompt takes.
03 · Grade
Checked against the truth
Every answer (each model, a plain majority vote, and the consensus) is scored against the dataset's known-correct option.
The headline number
Accuracy across all 314 questions.
The consensus comes out on top, but only just. At the top of the field the gaps are tiny, and the 95% confidence intervals (the whiskers) overlap heavily. Read honestly, the leading systems are in a statistical tie. The takeaway isn't "we win every question." It's that the consensus reliably tracks the best models instead of dragging toward the average.
Overall accuracy on MMLU-Pro
Share of 314 questions answered correctly · whiskers show the 95% confidence interval
Accuracy (%)
100
75
50
25
0
consens.ioconsensus
ClaudeOpus 4.8
GPT-5.5ΣMajority
vote
vote
Gemini3.5 Flash
Grok 4.3DeepSeekV4 Pro
MistralMedium 3.5
consens.io consensus
Single model
Majority vote (baseline)
95% confidence interval
Where it gets interesting
When the models disagree, the consensus pulls clear.
On most questions the models already agree, so combining them changes little. The real test is the 68 questions where they split. Here picking the wrong model costs you, and the gap opens up. The consensus lands on top again, and well above a naive majority vote. The whiskers are wide (only 68 questions), so treat this as a strong signal, not a settled fact.
Accuracy on the 68 disagreement questions
Only questions where the six models did not unanimously agree · whiskers show the 95% confidence interval
Accuracy (%)
100
75
50
25
0
consens.ioconsensus
ClaudeOpus 4.8
GPT-5.5ΣMajority
vote
vote
Gemini3.5 Flash
Grok 4.3DeepSeekV4 Pro
MistralMedium 3.5
consens.io consensus
Single model
Majority vote (baseline)
95% confidence interval
What it means
Three honest reads.
#1The consensus ranks first overall and on the hard questions, though the top of the field is within the margin of error.
+32ptsOn disagreement questions the consensus (72.1%) sits far above the weakest model (39.7%): picking wrong is expensive.
0Errors and abstentions for the consensus: it produced a clean, parsable answer on every single question.
Never the weak link
It tracks the best, not the average
Fusing six answers doesn't regress to the mean. The consensus stays at the top of the pack instead of being pulled down by weaker models.
Better than a vote
Reading answers beats counting them
The consensus reasons over the six answers and beats a plain majority vote, clearest exactly when the models split.
No overclaiming
The gaps at the top aren't significant
With 314 questions the confidence intervals overlap. We're not claiming a knockout, just that combining beats betting on one model.
The fine print
Exactly how this was measured.
Setup
- Dataset: 314 questions from MMLU-Pro, an expert-level multiple-choice benchmark, pooled across three runs from different sampling strategies (disagreement-enriched, category-balanced, and random).
- Closed-book: web search, retrieval and tools were disabled for every model. Each answered from its own knowledge only.
- Models: GPT-5.5, Claude Opus 4.8, Gemini 3.5 Flash, Mistral Medium 3.5, DeepSeek V4 Pro and Grok 4.3, the same six the app uses.
- Consensus: built with the app's production consensus logic; the synthesizing model for this run was Gemini 3.5 Flash.
- Grading: one attempt per question, scored against the dataset's ground-truth option. Majority vote is a simple plurality of the six model answers.
Reading the numbers
- Whiskers are 95% confidence intervals. Where they overlap, the difference is not statistically significant, and at the top of the overall chart, they overlap a lot.
- The disagreement subset is only 68 questions, so its intervals are wide. The ordering is a strong, consistent signal rather than a proven result.
- This is a snapshot of one frozen prompt and model line-up. We're keeping the prompt fixed so the run keeps collecting comparable, unbiased data over time.
Try it yourself