Benchmark sources
Start with the website that publishes the board
Each source below is a publisher. Open one to switch between the leaderboards they host — instead of treating every board as a separate site section.
cursor.com
IDE agent evalsCursor
Cursor’s first-party agent evals for IDE-native multi-file work. CursorBench is vendor-run from real Cursor sessions — useful inside Cursor, not a public reproducible harness.
3 boards · 63 local rows
arena.ai
Arena
Live human-preference and agent boards from Arena (LMArena). Scores are board-specific — Text Arena Score, WebDev Arena Score, and Agent net improvement are not interchangeable. The Arena product home at arena.ai/ and the long-running lmarena.ai leaderboards are the same operator (verified 2026-08-29 via arena.ai root and the LMArena research pages; see https://lmarena.ai/ for the historic domain).
3 boards · 596 local rows
artificialanalysis.ai
Artificial Analysis
Independent evaluations across composite intelligence, hard reasoning, science reasoning, and agentic terminal tasks. Multiple boards live under one publisher.
5 boards · 127 local rows
deepswe.datacurve.ai
DeepSWE
Long-horizon engineering evaluation that publishes pass rate together with cost, output tokens, and agent steps.
1 board · 28 local rows
swebench.com
SWE-bench
Repository-level GitHub issue resolution benchmarks. Verified (built by OpenAI with the SWE-bench team) is a human-filtered historical subset — saturated, with no official runs after February 2026.
1 board · 47 local rows
livebench.ai
LiveBench
A contamination-aware suite that refreshes questions over time. Category and subtask views live on the source site.
1 board · 58 local rows
Model providers
Provider catalogs
System limits from provider documentation — context capacity is not a quality benchmark.
1 board
A better comparison habit
Use a board to create a shortlist, then test a task.
A publisher can host several boards. Stay on one source page and switch boards instead of mixing Arena Score with DeepSWE pass rates — or CursorBench correctness — as if they were the same unit.
Trust signal
Named publisher. Named board. Measurement date.
Open the source website when you need the full live table. Some pages keep a full local mirror of a board; others keep a readable local shortlist plus the caveats that travel with it. Do not treat every source card as a shortlist.
Example live write-up: CursorBench