LLM Leaderboard · Metacritic for Models

One page to pick the right model for the job: public quality evals as visible per-source bars, per-million-token pricing, refusal-rate data by domain, and sliders for what you care about. Move a slider, the ranking re-computes live.

Data as of loading – models tracked Refreshed daily

Quality: LMArena leaderboard (CC-BY-4.0). Pricing and context: OpenRouter public API (prices include OpenRouter's service fee, so they sit slightly above direct vendor prices). Refusal rates: moderationbias.com independent audit, LLM-judged, used as a labeled proxy. Param counts: Hugging Face Hub API.

Quality vs price 70% quality

Max price / 1M output tokens $200

Minimum context window

Deployment

Local quantization

VRAM cap (local models) 160 GB

Refusal ceiling (selected domains) off

Models with no refusal data are never excluded by this slider.

#	Model	Score	Arena	Coding	Math	Price / 1M	Context	Deploy	Why

Methodology

The bars

Arena: LMArena crowd-preference Elo from the official leaderboard catalog, rescaled to 0-100 for the bar. The number shown is the raw Elo.
Coding / Math: LMArena category Elo for the same model, same source, same scale. Missing categories show as n/a, never as zero.
Score (the big number): w x quality + (1-w) x value, where quality = clamp((elo - 1200) / 300) and value = 1 - clamp((log10(out_price) - log10(0.05)) / (log10(200) - log10(0.05))). The weight w is your Quality-vs-price slider. Models with no public price are scored on quality alone and flagged "price unknown".

Refusal data

Refusal rates come from an independent moderation audit that prompts each model identically at temperature 0 and has an LLM judge classify refusals. It is a moderation-bias study, not a safety benchmark, so we present it as a labeled proxy with its source linked, and we never filter out models that lack coverage.

VRAM estimates

For open-weight models with verified parameter counts: VRAM = params x bytes-per-param x 1.15 (FP16 = 2 bytes, INT8 = 1, INT4 = 0.5; the 1.15 covers KV cache and runtime overhead). Rough estimate, not a deployment guarantee.

Honest caveats

Arena scores measure crowd preference, not task accuracy. They skew English, skew toward heavily-voted proprietary models, and update when LMArena publishes, roughly weekly. A model can be the best at your task while sitting mid-table here.
Prices move. OpenRouter's feed adds its own service fee on top of provider prices, and vendors change list prices without notice. Treat every price as "as of the data date", not a quote.
Refusal coverage is thin (a dozen or so of the 80 models). Absence of data is not evidence of low refusal.
VRAM math ignores context-length KV growth, batching, and framework overhead. Real deployments vary by 20-40%.
Model names shift across sources; matches are by normalized name and spot checked, but a mismatch can slip through. The per-model source links let you verify.