By Model VS · Guide text updated . Essay text is ours. Measurements belong to the named sources and are inserted from the JSON files listed below. Send a correction if a table does not match its file.
Start with the question you actually have.
How to run a private model bake-off on your own tasks
After the public names exist, freeze the setup, plan the sample, keep failures, and allow an inconclusive result.
Latency and interactivity vs benchmark quality
A high exam score can still feel slow. Measure time-to-first-token, throughput, and thinking tokens on your own calls, and keep them off the quality cell.
Cost-per-quality triage
The snapshots have no prices. Copy one quality field, and fill the public list price yourself.
Provider card versus the leaderboard shortlist
A card can cite what you read on that page. A rank or a benchmark value still needs the saved shortlist.
Vote count versus interval width
A larger vote total is not a tighter band and not a win over the next rank. The rows come from the saved Arena file.
When Arena and LiveBench disagree
Exact source ids that reverse order. Follow the job. Do not average a preference rating with a task score.
Reading specialist and HF cards
The row stores a metric, an id, a flag, and a link. The card is not an endorsement, and the README is not copied here.
A 30-minute model shortlist
One acceptance sentence, one category, at most three ids. The worked column is a fictional example.
How to read LM Arena intervals
Preference votes, overlapping bounds, and why a high or thinking suffix is a different saved record.
LiveBench categories for shortlisting
Match a task to a category. Overall is an unweighted mean and can hide the column you needed.
SWE-bench flags
The Verified subset and the submission flag are different facts. A percent without a harness is an incomplete citation.
October 2026 snapshot changelog
Entrants, exits, and rank moves between the 24 September archive and the current files, kept apart from editorial edits.
The rankings at /#leaderboard and the compare tool at /#compare are the interactive surfaces. How we rate is the protocol for testing a shortlist on your own cases. These essays sit between them: they say what a public number can support before you spend time on a pilot.
The files these essays were built against.
- LM Arena snapshot: leaderboard publish date 2026-10-02, saved October 6, 2026 (2026-10-06T09:00:53.013Z), 100 overall rows, dataset lmarena-ai/leaderboard-dataset, license CC BY 4.0.
- LiveBench snapshot: release 2026-06-25, saved October 6, 2026 (2026-10-06T09:00:12.143Z).
- Specialist snapshot: saved October 6, 2026 (2026-10-06T09:00:11.687Z); SWE-bench Verified dataset timestamp August 16, 2026; GPQA dataset timestamp September 21, 2026.
If those times are older than the upstream sites, the essays are still about the files, not about a board we did not copy. The changelog keeps the September archive address so a later sync does not erase the comparison.
Radar and Skills are indexes. They are not these essays.
AI Radar lists public projects, releases, papers, datasets, and MCP records from named sources. Skills lists public GitHub repositories that contain a valid skill file. Both pages are reachable, both say they are excluded from search indexing, and both are omitted from the sitemap on purpose. A star count or a recent push is a discovery signal. It is not a model evaluation, and a third-party script in a skill file is never run by this site.
The old addresses /rankings, /compare, and /radar redirect to the leaderboard, the compare anchor, and the Radar index. Use those destinations. Do not treat a redirect as a second copy of the essay library.