Use the API or connect an MCP client
Read final BenchAlign scores and ranks for Overall, Coding, Agentic, and Knowledge. Query all admitted current benchmark results in their native units. Search models by name or creator, and benchmarks by name, category, or tag to resolve their IDs.
Free includes 250 monthly reads of current data. Pro is $49 USD/month for 25,000 reads and three months of history; Research is $249 USD/month for 100,000 reads and all retained eligible history.
One Data key for REST and MCP
- REST base URL
https://data.benchlm.ai/v1- MCP endpoint
https://data.benchlm.ai/mcp
Choose your history window
Free
$0
250 monthly reads · 30 requests/minute
Current data only.
Create a free accountPro
$49 USD/month
25,000 monthly reads · 60 requests/minute
Three rolling calendar months of available history, measured by UTC date.
See Pro accessResearch
$249 USD/month
100,000 monthly reads · 60 requests/minute
All retained eligible history with no age cap. Coverage varies by source, model, and date.
See Research accessREST and MCP share one allowance. Check retained coverage in your account; checkout opens when available.
What you can query
Current rankings
Final BenchAlign scores, published ranks, evidence labels, score intervals, and the scoring date and methodology version. Available on all plans.
Source benchmark results
Current results from admitted source boards, with native scores, exact model profiles, and source notices. Free includes these results. Use the coverage query to see available boards, row counts, metrics, and licenses.
Model and benchmark catalogs
Names, IDs, creators or categories, source references, and sourced model release dates. A benchmark’s presence in the catalog does not mean its numerical results are available through the service.
Recorded history
Pro and Research queries cover saved BenchAlign rankings, publisher snapshots, and eligible benchmark measurements and catalog changes where retained. Check historical coverage for the model, benchmark, and period you need.
Source results retain exact model profiles and native units. Run dates come from the publisher; archive dates describe saved states. BullshitBench v2 has one retained snapshot. Source-run scores do not establish historical leaderboard positions. Publisher snapshot history preserves ranks when the source reported them. Check coverage before requesting a period.
Commercial display or redistribution? Request a license
Free, Pro, and Research are access plans for queries and internal analysis. Publishing BenchLM-owned outputs or the compiled feed in a customer-facing commercial website, app, or product needs a separately agreed display license. Redistributing that material through an API, downloads, or resale needs separately agreed scope. Paid plans add reads and available history.
Individual quotations and attributed embeds keep their existing permissions. Independently licensed publisher measurements keep their own commercial-use and redistribution rights. Retain attribution, source notices, and licenses. Read the Data access terms or request a commercial license.
Download the website’s JSON files
The site has 1003 public model profiles and 717 canonical benchmark pages as of October 29, 2026. These eight static downloads contain selected model, benchmark, ranking, price, and runtime fields where available. Source-specific benchmark definitions can share a canonical page; the files do not include every page or source result.
leaderboard.jsonPublished overall and category rankings, scores, evidence labels, and score intervals.
Schema 1.0 · Source last updated: October 29, 2026
models.jsonModel metadata, available benchmark scores, context windows, release metadata, and provenance.
Schema 1.0 · Source last updated: October 29, 2026
benchmarks.jsonBenchmark definitions, category weights, and per-benchmark metadata.
Schema 1.0 · Source last updated: October 29, 2026
pricing.jsonAvailable API prices per million input and output tokens.
Schema 1.0 · Source last updated: October 29, 2026
speed.jsonRuntime metrics: tokens per second and first-answer latency.
Schema 1.0 · Source last updated: 2026-10-28
comparisons.jsonPrecomputed head-to-head model comparison data.
Schema 1.0 · Source last updated: October 29, 2026
stats.jsonDated statistics with self-contained statements and links to the source pages and individual facts.
Schema 1.0 · Source last updated: October 29, 2026
price-index.jsonMonthly token price index series, tier sub-indexes, and current constituent prices.
Schema 1.0 · Source last updated: October 29, 2026
Download manifest.json for the eight static files’ content versions, schema versions, byte sizes, and SHA-256 hashes. The ninth dataset download, ai-model-changes.json, is a dynamic endpoint for confirmed model releases and retirements. Its response is excluded from the manifest’s static-file hashes.
Record the content version and source dates
Each static file’s dataVersion is derived from its content using SHA-256. The version calculation leaves out the root generatedAt clock and the version field itself; nested source dates remain part of the content. The manifest’s dataVersion identifies the complete static download set.
The manifest’s sha256 hashes each file’s exact downloaded bytes. Regeneration can change those byte hashes when the clock changes, even if the content version stays the same.
generatedAt records when a file was built. sourceLastUpdated is the freshness marker for its underlying data. Neither field establishes when a benchmark was evaluated. Use the individual result’s source and provenance dates where supplied; coverage and evaluation dates differ by source. Website data was last synchronized on October 29, 2026.
Use the published score fields
In models.json, top-level displayScore and overallRank describe the published overall leaderboard selected by publicRankingMode. The nested scores.displayScore, scores.overallScore, and scores.rawOverallScore fields preserve compatibility values, and provisionalDisplayScore is provisional. Those fields can differ from the published board and should not be used to reconstruct it.
In leaderboard.json, items[].score and categories[].score are published scores for each declared rankingMode. Keep that mode with the score when comparing or citing rankings.
For REST and MCP queries, see the separate query methods and coverage.
Website download licensing and citation
The website JSON downloads and benchlm.ai/api/data exports are published under CC BY-NC 4.0. Anyone may quote individual scores with attribution. You may reproduce tables and charts in articles, papers, and research for non-commercial use, with attribution. Commercial use of these exports requires a license from us; get in touch. Underlying publisher licenses and notices continue to apply.
Copy a dataset citation
Cite the dataset version you downloaded and replace YYYY-MM-DD with your retrieval date. The citation year below is the export year, not the year of every benchmark result. Save the files and manifest used in your analysis; the download URLs serve the current exports.
@misc{benchlm_dataset_2026,
author = {{BenchLM.ai}},
title = {{BenchLM LLM Benchmark Dataset}},
year = {2026},
url = {https://benchlm.ai/data},
version = {sha256:197e07ee2463cab190409fff27787414146d3447b49513bbd969b83e32f147e5},
note = {Retrieved: YYYY-MM-DD}
}See how scores are built and benchmark confidence for provenance.
Feeds, crawler files, and embeds
llms.txt and llms-full.txt provide crawler references. Follow RSS or Atom, or read data-change events. The embeddable leaderboard stays current and includes attribution.