LLM Benchmarks & Evaluation Harnesses
Model rankings shift every few weeks. Rather than embed a scoreboard that goes stale, this page links straight to the leaderboards and harnesses the industry actually uses — plus a live feed of the latest model news.
Top Benchmark Leaderboards
Four widely-cited, independently-run leaderboards, each measuring something different — human preference, price/performance, or broad multi-metric quality.
LMArena
Human preference · Elo ranking
Blind, pairwise human voting between two anonymous models, aggregated into an Elo-style leaderboard across chat, coding, and vision arenas.
Artificial Analysis
Quality · price · speed
Independent benchmarking across API providers on intelligence, cost per token, and latency/throughput — useful for model-routing and procurement decisions.
Stanford HELM
Holistic multi-metric evaluation
Academic, reproducible evaluation across accuracy, calibration, robustness, fairness, and efficiency — not just a single leaderboard score.
Hugging Face Open LLM Leaderboard
Open-weight model benchmark suite
Standardized automated benchmarks for openly-licensed models. Hugging Face has retired and relaunched this project before — check it's still active before citing it.
Evaluation Harnesses
Leaderboards tell you where models rank. Harnesses are the actual tooling to run those evaluations yourself, against your own prompts and data.
EleutherAI lm-evaluation-harness
The open-source harness behind most public leaderboards — hundreds of standardized tasks, pluggable against any model API.
OpenAI Evals
A framework for writing custom, repeatable evals against your own tasks rather than only generic benchmarks.
HELM (the harness)
The same codebase that produces the Stanford HELM leaderboard above — runnable against your own models.
Already in this map's Capability Directory: Braintrust, DeepEval, Patronus AI, and Ragas are the production-grade eval platforms referenced under AI Evaluation & Quality Assurance.
Latest Model News
Live-fetched on page load from the Hugging Face Blog RSS feed — one of the highest-signal public feeds for new model releases.
Loading latest items…
Fetched client-side via a public RSS-to-JSON proxy — no server on this site, so availability isn't guaranteed. If it's down, use the direct link shown in its place.