There are dozens of LLM leaderboards and no way to read them together.
GPQA reports a fraction. Arena reports Elo near 1500. Vending-Bench reports dollars, and its range runs from −31 to +10,940. One board says Claude wins, the next says GPT, a third ranks by tokens routed. None of them know whether you are writing code, running agents, or trying to spend under a cent per million tokens.
So people pick a model from whichever chart they saw last.
Pulls 53 benchmarks, 6 human-preference arenas and live pricing into one comparable index — then lets you re-weight it for the job you actually have.
708 models · 2,829 measurements · 305 priced · 3 independent sources
Tell it what you need. Eight task presets — coding, autonomous agents, research, maths, chat, images, cheap-at-scale — each re-weights the categories and re-ranks all 708 models instantly, in your browser. Or move the sliders yourself.
It tells you how much to trust the answer. A model measured once does not outrank one measured thirty times. Every score carries a confidence, and the index is shrunk toward the median in proportion to how thin the evidence is.
See what quality costs. The price/quality frontier shows which models nothing cheaper can match — and which you are simply overpaying for.
And a few things other boards don't show you:
- 📈 Frontier over time — the exact moment each capability ceiling moved, and who moved it
- 🌍 Who builds models — the US/China race, by median index rather than headline count
- 🎭 Humans vs benchmarks — where blind human votes disagree with test scores, and by how much
- 🔬 Raw benchmark explorer — every leaderboard unnormalised, if you distrust the maths
- ⚖️ Side-by-side — up to four models across all eight categories
Filter by vendor, country, open weights, price cap, context length, tool calling, vision, reasoning mode or free tier. Every view is a shareable URL. Russian and English. Light and dark.
Four steps. All of them are in scripts/, and steps 2–4 run again
in your browser every time you touch a weight.
1 · Normalise. Each benchmark is stretched to 0–100 against its own cohort of current models, clamped at the 2nd and 98th percentile. This is what makes a 0–1 fraction, an Elo rating and an unbounded dollar score comparable at all.
2 · Roll up. Normalised results are weighted into eight capability categories: reasoning, maths, coding, agentic, multimodal, long context, writing, human preference.
3 · Adjust for evidence. Confidence is the geometric mean of breadth (how much of the weighted category space is covered) and depth (total measurements). Neither can compensate for the other — 100% breadth with one datapoint each is still thin. The index is pulled toward the median in proportion to what's missing.
Without this step, a model with two arena rows outranked one with thirty-three measurements across five categories. It also means a brand-new release starts out looking modest — which is honest. There genuinely is less evidence about it.
4 · Re-weight. Presets and sliders rerun steps 2–4 client-side from the identical formula. No server, no round-trip.
Epoch's own composite (ECI) sits in its own column rather than being folded in. It draws on the same underlying runs, so including it would double-count — but it makes a useful second opinion, and you can sort by it.
The same model arrives as claude-opus-5_max from Epoch,
claude-opus-4-6-thinking from LMArena, and anthropic/claude-opus-4.8 from
OpenRouter.
Resolution anchors on OpenRouter's catalogue — the only source listing products
you can actually buy — and matches foreign names by progressively stripping
reasoning-effort, release-channel and build suffixes until something lands.
Order matters: qwen3.7-max hits on the first try and keeps its max (a
product tier), while gpt-5.5-high misses, drops high, and correctly
resolves to gpt-5.5.
Folding . into - is what makes it work at all. Without it,
claude-opus-4.8 (a price, no scores) and claude-opus-4-8 (34 measurements,
no price) stay two different models forever.
| Source | Provides | Licence |
|---|---|---|
| Epoch AI Benchmarking Hub | 53 benchmarks + the Epoch Capabilities Index | CC BY 4.0 |
| LMArena via the official HF dataset | Blind pairwise human votes as Elo, 6 arenas | CC BY 4.0 |
| OpenRouter | Pricing, context windows, modalities, capabilities | Public API |
A source only earns a place here if its terms allow republishing what it provides. Artificial Analysis was evaluated and left out for exactly that reason: its free Data API tier is "internal use only; no redistribution", and this site is public. The three above cover every part of the index on their own.
Benchmarks do not measure tone, production reliability, API quality, rate limits or speed. Different tests run with different scaffolds and reasoning settings, so any cross-test comparison is approximate. Prices are OpenRouter's and may differ from a direct vendor contract. Open weights does not mean free — count the hardware.
No framework. No charting library. Charts are hand-drawn SVG.
| gzipped | |
|---|---|
| JavaScript | 16.3 KB |
| CSS | 4.3 KB |
| HTML | 1.1 KB |
| Data (meta + models) | 40 KB |
| First load | ≈ 61 KB |
Per-model score breakdowns live in a second payload fetched when the browser goes idle — that split alone cut first load by 23 KB. Everything is computed at build time, so the browser only ever filters and sorts arrays it already holds: no API calls, no CORS, and nothing breaks when an upstream source has a bad day.
npm install
npm run collect # fetch + normalise all three sources (~10 s)
npm run devpublic/data/ is generated, not committed.
| Command | Does |
|---|---|
npm run collect |
Rebuild the data layer from upstream |
npm run collect -- --offline |
Rebuild from .cache/ — no network, instant |
npm run typecheck |
tsc --noEmit |
npm run build |
Production bundle into dist/ |
Adding a benchmark is one row in scripts/registry.mjs:
its id, its CSV column, a category and a weight. Nothing else changes — it
appears in the index, the explorer and the model sheets automatically.
GitHub Actions builds and publishes to GitHub Pages on every push to main,
weekly on Monday at 05:25 UTC, and on manual dispatch. Pull requests run the
same build without deploying. The custom domain comes from public/CNAME.
The collector caches every upstream response and falls back to the last good copy on failure, and CI restores that cache between runs — a source outage degrades one panel instead of publishing a leaderboard with a hole in it.
First-time setup on a fresh fork:
- Settings → Pages → Build and deployment → Source: GitHub Actions.
- Point DNS at GitHub Pages — a
CNAMErecord for the subdomain to<owner>.github.io.
No secrets or API keys are required. Every source is public.
Code is MIT. Aggregated data belongs to the sources above and stays under their terms — CC BY 4.0 for Epoch AI and LMArena, which requires attribution when you reuse it.