AI model benchmarks
AI benchmarks, tracked and explained
New AI models launch every few weeks, each with a chart claiming it is the best. We track independent benchmarks, chart who actually leads, and explain what the scores mean for the tools you pick for your business.
Benchmarks we track
Pick a benchmark
Each page shows the current ranking, the trend over time, a head-to-head comparison and a plain-English method note.
Epoch Capabilities Index
LiveOne score for general AI capability, combined from dozens of benchmarks. The best single answer to "which model is strongest?"
Software engineering
Coming soonCan a model fix real bugs in real codebases? The benchmark most teams look at before choosing a coding assistant.
Expert reasoning
Coming soonGraduate-level science questions written to resist search. A measure of careful reasoning on hard material.
Advanced math
Coming soonResearch-level math problems that take experts hours. Shows where the hardest reasoning still breaks down.
Model personality
Coming soonHow models communicate: direct or wordy, warm or neutral, confident or hedged. It matters for anything customer-facing.
Cost and speed
Coming soonCapability per dollar. The model at the top is rarely the one you should run on every task.
How to read benchmarks
Four rules before you trust a chart
Check who ran it
Vendor-run numbers use the vendor's best settings. We use independent evaluators that test every model the same way.
Look at the error bars
A one or two point lead is often noise. If the ranges overlap, call it a tie.
Match the test to the task
A math benchmark says little about customer email. Read the benchmark that looks like your work.
Test on your own data
Benchmarks build the shortlist. A week of trials on your real tasks makes the decision.
Free AI audit
Want the shortlist tested on your workflows?
We pick the models that fit, run them on your real tasks and show you what to automate first.