Compare exact models
Choose a lab and two models. The Epoch Capabilities Index (ECI) combines benchmark scores on a common point scale; it is not an accuracy percentage.
Capability over time
A dashed connector shows only the chosen comparison; it does not imply a generational lineage.
View plotted data
What do the numbers mean?
Entry or release? An entry is an imported model group or catalog label. Aliases can overlap. Evaluation configurations are grouped, not counted as new generations.
ECI or accuracy? The Epoch Capabilities Index combines many benchmarks on an arbitrary point scale. A gain of 5 points is not “5% smarter”. A missing ECI does not mean a weak model.
Which date? Benchmark release dates take priority; catalog publication dates fill gaps. Date spacing describes these sources, not a complete launch calendar.
What changed in the benchmarks?
Raw benchmark scores on a common 0–100% scale. Hollow: baseline. Filled: comparison model. Best observed configuration per model; settings can differ.
View exact scores as a table
Changes are percentage points. Missing scores are absent from the imported benchmark export, not zero. The winning evaluation configuration is shown under each score. Settings and inference budgets may differ; these scores do not measure accuracy across an entire domain.
Models behind the comparison
Source-reported identity, dates, evaluation configurations and available metadata for the two selected models.
Explore the model inventory
All source entries are available, including models without ECI. API reasoning settings are grouped under the source model, not counted as new generations.
What does the recorded release rhythm look like?
One year ending on the import date, 16 Nov 2026. Dates combine benchmark release dates and catalog publication dates where no release date is available.
Publication dates
Models on the same date count as one event. These are recorded source dates, not a census of major generations.
Read the two signals together
A shorter gap means more frequent publications. A larger ECI gain means a larger change in the composite benchmark estimate. These are different signals; neither measures a lab’s scientific productivity.
Source: Epoch AI, CC BY 4.0. Benchmark export imported 16 Nov 2026 and combined with the existing notable-model catalog. ECI uses an arbitrary point scale: differences are points, not percentages. Error bars show source-provided 90% credible intervals where available; we do not estimate confidence in pairwise differences from those intervals. Source model groups and exact catalog labels are retained; aliases and overlapping release records may remain.