For you.

Your use cases are saved to your account. Manage your account.

Receive benchmark updates in your RSS reader, even when this page is closed.

Keep codes and feed links private. A new code or feed link replaces the previous one. Restoring a workspace replaces the workspace open in this browser.

Each capability combines several published benchmarks. We put each test on a fixed scale, then average the available results within that capability. Your priorities decide how much each capability contributes to your personal index.

A model needs at least two benchmark results in every selected capability to enter this leaderboard or its preview. Models with only one result in a selected capability are listed separately as unranked. Limited evidence in a capability you haven't selected doesn't affect this ranking. A missing capability still means no overall score. Coverage can differ between models. Open a model to see exactly which tests contributed, including the published settings. Missing tests are never invented or treated as a zero.

The comparison with your current model uses only tests both models completed, with at least two shared tests per capability. Upgrade updates require a higher personal index and a higher score on that shared evidence.

Index points aren't a success percentage. Small differences may be evaluation noise. These rankings help you choose what to try on your own work.

Normalization is fixed: 50 is a benchmark's reference mean, and 15 points is one reference standard deviation. Adding a model doesn't change that scale. Published scores can still be revised by their authors. Overlapping capabilities can give a test weight in more than one part of your mix.

Evidence comes from Epoch AI (CC BY 4.0), EQ-Bench, Lech Mazur, Roboflow (Apache 2.0), Andon Labs, Surge AI, Terminal-Bench, Terminal-Bench Science, WeirdML, LiveBench, ExtractBench, and AutomationBench. We use their published results, not tests on your private data.

Method cadence-personal-4. New reference scales frozen September 13, 2026. Evidence refreshes every two hours. Optional task suggestions select capabilities only; they never invent scores or choose models.