Reference leaderboards

Leaderboards we link to but do not republish. Licence or terms-of-service constraints prevent ingestion; we surface what each one measures and link out so you can read the source directly. Where part of a publisher's output reaches us under an open licence by another route, the card names those boards and links to their report cards.

CursorBench

Code

What it measures: Cursor's own benchmark: ambiguous multi-file tasks taken from real Cursor sessions, scored per model with cost, tokens and steps; 68 configurations, updated through October 2026.
When to consult it: Seeing how models behave inside one specific coding agent — bearing in mind the vendor chose the tasks and runs the harness.

Scale SEAL leaderboards

Mixed

What it measures: Scale AI's private, expert-written evaluations (SEAL) across coding, agents, multilingual and safety, with results kept on their site.
When to consult it: A second opinion from private test sets the labs cannot have trained on.

SWE-bench Pro (Scale)

SWE

What it measures: Scale's harder successor to SWE-bench Verified: 1,865 tasks from copyleft repositories, 731 public ones scored with ± intervals; no result file to ingest yet.
When to consult it: Checking a coding agent against tasks chosen to resist contamination.

METR time horizons

Agent

What it measures: METR's time-horizon estimate: the length of task, in human minutes, a model completes with 50% success, with intervals; only raw runs are published, the fit is theirs.
When to consult it: A measure of autonomy that no accuracy benchmark gives you.

Vals.ai benchmarks

Mixed

What it measures: Vals.ai's 30 independent benchmarks across legal, finance, healthcare, coding and math, each with accuracy, cost and latency.
When to consult it: Domain-specific checks (law, finance, medicine) the general boards do not cover.

Aider polyglot leaderboard

Code

What it measures: Aider's code-editing benchmark over 225 Exercism exercises in six languages, with the cost of each run; last updated November 2025.
When to consult it: Historical reference for editing-format reliability; the board has gone quiet.

Artificial Analysis

Mixed

What it measures: Inference speed, latency, throughput, cost, and a composite quality index for hosted LLM endpoints.
When to consult it: Picking which API to call for a latency- or cost-sensitive workload.

LLM Stats

Mixed

What it measures: Cross-benchmark aggregator (50+ benchmarks incl. GPQA, MMLU-Pro, AIME, MATH, HumanEval, MMMU) with pricing, context window, and provider availability per model. Their ToS permits redistribution with attribution — we're not blocked by license, just haven't wired up their API yet.
When to consult it: Comparing pricing and feature support across providers in one view, or checking a benchmark axis (math, vision, general knowledge) we don't ingest directly yet.

BenchLM

Mixed

What it measures: A cross-benchmark "bench-align" index over agentic, coding, reasoning, knowledge and instruction-following.
When to consult it: A quick cross-check on our own rankings. Read the per-category numbers rather than the headline index — a third of its models are scored by estimation, not measurement.

Open LLM Leaderboard (HuggingFace, archived)

Mixed

What it measures: Automated evals (IFEval, BBH, MATH, GPQA, MuSR, MMLU-Pro) over open-weight models.
When to consult it: Historical comparison of 2024-era open-weight models. Frozen since March 2025 — it contains no frontier models.

OpenRouter Rankings

Mixed

What it measures: Token and request share across OpenRouter's own traffic, plus third-party benchmark scores it republishes.
When to consult it: Seeing what people actually run in production, and at what cost. Read it as adoption, not capability — a model can top it by being cheap and default.

Vellum LLM Leaderboard

Mixed

What it measures: Curated multi-benchmark roundup with short editorial notes per model.
When to consult it: A quick "what's hot" pass when you only have a minute.

Onyx — Best LLMs

Mixed

What it measures: Editorial best-of list across coding, reasoning, and math benchmarks.
When to consult it: A second opinion against the headline benchmarks.

DeepSWE

SWE

What it measures: Open-weights agent's reported SWE-bench Verified score, surfaced in its own GitHub README.
When to consult it: Tracking the open-weights state of the art on real-repo software tasks.

Vending-Bench

Agent

What it measures: Long-horizon agent coherence — managing a simulated vending business over 20M+ tokens per run.
When to consult it: Stress-testing whether an agent can sustain coherent action over hours, not turns.

MathArena

Math

What it measures: Frontier model performance on math olympiads (AIME, IMO, HMMT) on contamination-free problems.
When to consult it: Tracking how reasoning models do on the hardest contest math available.

LiveCodeBench

Code

What it measures: Contamination-free coding benchmark sourced from time-windowed competitive-programming problems.
When to consult it: A coding-axis sanity check that resists training-data leakage.

Terminal-Bench

Agent

What it measures: Agentic evaluation of models operating a real terminal — task success across a curated harness of shell / dev-loop scenarios.
When to consult it: One of the most actively-updated agentic benchmarks in 2026; check the site for the current cross-vendor ranking.

Stanford HELM Capabilities

Mixed

What it measures: Stanford HELM's Capabilities project — holistic per-scenario evaluation of frontier models across dozens of tasks.
When to consult it: Authoritative multi-axis reference from Stanford CRFM. Latest release v1.15.0 dates from 2025-11-24.

Arena-Hard-Auto

Chat

What it measures: Automated hard-prompts benchmark from LMSys — an LLM-judge variant complementary to the live Chatbot Arena.
When to consult it: Reproducible hard-prompt comparison when human-preference Elo (Arena) is not what you need.