routerflow.artthe marketour probecomparemethodologyprobestaxonomyprober methodapi/v12026-11-10

The data, as tables

Everything routerflow collects and measures, as tidy tables you can query: one Parquet file per table, and one DuckDB file holding them all. Rebuilt every hour on the collection server from the daily snapshots, the raw files and the probe records. No key, no account.

Base URL: http://167.71.161.149/data/tables/ (a directory listing). Dictionary as JSON: http://167.71.161.149/data/tables/tables.json. License: CC BY 4.0; upstream licenses apply per source.

Query it

DuckDB, straight off the URL, nothing to download:

INSTALL httpfs; LOAD httpfs;
SELECT route, count(*) AS requests, round(sum(billed_usd), 2) AS usd
FROM 'http://167.71.161.149/data/tables/probe_request.parquet' GROUP BY 1;

pandas:

import pandas as pd
df = pd.read_parquet('http://167.71.161.149/data/tables/market_daily.parquet')

Everything at once: download http://167.71.161.149/data/tables/routerflow.duckdb and open it with duckdb routerflow.duckdb.

A version to cite: the files here are rebuilt every hour. Monthly frozen copies live under http://167.71.161.149/data/tables/releases/<date>/, each with a MANIFEST.json of sizes and hashes, listed in http://167.71.161.149/data/tables/releases/index.json; see RELEASES.md.

A cache pair is two rows of probe_request with the same session_id: repeat = 1 (cold) and repeat = 2 (the same bytes again). What the provider recognized:

SELECT a.route, a.model, a.provider, a.endpoint, b.cached_tokens * 1.0 / b.prompt_tokens AS recognized, b.billed_usd / a.billed_usd AS second_over_first
FROM 'http://167.71.161.149/data/tables/probe_request.parquet' a JOIN 'http://167.71.161.149/data/tables/probe_request.parquet' b USING (session_id)
WHERE a.workload = 'cache-pair' AND a.repeat = 1 AND b.repeat = 2 AND NOT b.failed;

One provider across routers: provider is the OpenRouter provider slug in every table, so what a provider lists, what others measured and what we measured join on it:

SELECT q.provider, q.endpoint, q.prompt_usd_per_m AS listed, s.effective_prompt_usd_per_m AS effective, s.cache_hit_rate_1m, t.ttft_p50_ms AS trustedrouter_ttft, median(p.ttft_ms) AS our_ttft
FROM 'http://167.71.161.149/data/tables/openrouter_endpoint_quote.parquet' q
LEFT JOIN 'http://167.71.161.149/data/tables/openrouter_endpoint_stats.parquet' s ON s.snapshot = q.snapshot AND s.model = q.model AND s.endpoint = q.endpoint
LEFT JOIN 'http://167.71.161.149/data/tables/trustedrouter_provider.parquet' t ON t.snapshot = q.snapshot AND t.provider = q.provider
LEFT JOIN 'http://167.71.161.149/data/tables/probe_request.parquet' p ON p.endpoint = q.endpoint AND p.model = q.model AND NOT p.failed
WHERE q.model = 'z-ai/glm-5.3' AND q.snapshot = (SELECT max(snapshot) FROM 'http://167.71.161.149/data/tables/openrouter_endpoint_quote.parquet') GROUP BY ALL ORDER BY listed;

The leaderboards (schema 1.2)

Every public model leaderboard we read, in one long table: usage shares from routers, arena ratings, benchmark scores, list prices, measured speeds, uptime. One row per board, period, entity and metric (board_entry), with source and board saying who publishes it under which license, board_metric giving each metric's unit and direction, and the entity tables (model, lab, provider, app, ...) and alias tables giving one id per thing. Public rows only: a board is here when its redistribution is yes or attribution; unclear and no boards are collected into a private copy that is not served, and board lists them with public = false.

board_entry is one Parquet file per source and year, board_entry/source=<source_id>/year=<yyyy>/part.parquet, listed in http://167.71.161.149/data/tables/board_entry_files.json (path, rows, bytes, period range). Over HTTP DuckDB cannot glob a directory, so read the list:

import duckdb, json, urllib.request
base = 'http://167.71.161.149/data/tables'
files = [base + '/' + f['path'] for f in json.load(urllib.request.urlopen(base + '/board_entry_files.json'))['files']]
con = duckdb.connect(); con.execute('INSTALL httpfs; LOAD httpfs')
con.read_parquet(files).create_view('board_entry')
con.execute("SELECT source_id, count(*) FROM board_entry GROUP BY 1 ORDER BY 2 DESC").fetchall()

Or download http://167.71.161.149/data/tables/routerflow.duckdb: it holds the public board_entry as a table and the views board_latest, board_current and model_everywhere. The smaller files: board_entry_latest.csv.gz (what every board shows now), and every dimension, entity and alias table as <table>.csv.gz next to its Parquet.

Snapshots. Most sources publish the same history again every night, so board_entry keeps one row per key for each run of equal values: snapshot (= first_seen) is the day a value was first collected and last_seen the newest day it was collected unchanged. A key gets a new row only when its value, interval, rank, count or cost changes; board_latest is the newest row per key. Vercel revises its past days: its share for GLM 5.3 Flash on 2026-09-19 was read as 3.76 on the 19th (the day still running, last_seen 2026-09-19) and as 4.06 from the 20th on (a second row, snapshot 2026-09-20), so a query for what Vercel says now uses board_latest, and one for what it said on a night filters snapshot <= night AND last_seen >= night.

One model across every board (model_everywhere: one row per mapped model, board and metric at its newest period):

SELECT board_id, metric, rank, value, unit, period_end, source, tier
FROM model_everywhere WHERE model_id = 'z-ai/glm-5.3-flash' ORDER BY source_id, board_id, metric;

Usage share against an arena rating over time (a router's daily share joined to the arena board that was current on each day):

WITH usage AS (SELECT period_end AS day, value AS share_pct FROM board_latest
               WHERE board_id = 'vercel.leaderboard.models.day' AND metric = 'tokens' AND entity_id = 'z-ai/glm-5.3-flash'),
     arena AS (SELECT period_end AS published, value AS rating, rank FROM board_latest
               WHERE board_id = 'lmarena.text.overall' AND metric = 'rating' AND entity_id = 'z-ai/glm-5.3-flash')
SELECT u.day, u.share_pct, a.published AS arena_board_of, a.rating, a.rank
FROM usage u ASOF LEFT JOIN arena a ON a.published <= u.day ORDER BY u.day DESC;

The same benchmark as run by three sites (AIME: Epoch AI's own runs, MathArena and Vals; only Epoch's rows are public today, so on the served files this returns the Epoch column alone, and all three on a full build):

WITH aime AS (SELECT * FROM board_entry WHERE entity_type = 'model' AND entity_id IS NOT NULL AND lower(board_id) LIKE '%aime%' AND unit = 'pct'),
     latest AS (SELECT * FROM aime QUALIFY row_number() OVER (PARTITION BY board_id, entity_id ORDER BY period_end DESC NULLS LAST, last_seen DESC, value DESC) = 1)
SELECT entity_id AS model, source_id, board_id, effort, round(value, 1) AS accuracy_pct, period_end FROM latest
WHERE entity_id IN (SELECT entity_id FROM latest GROUP BY 1 HAVING count(DISTINCT source_id) >= 3) ORDER BY model, source_id;

Names: entity_id is the canonical id (the OpenRouter slug for models) resolved from data/curated/aliases/ when the tables are built, so a corrected alias reaches all history; entity_raw is what the source printed and is never edited. name_map shows how every name resolved, or why it did not. Thinking mode, agent scaffold and free or batch tiers are columns of their own (effort, agent, variant), never part of the id.

Read this first

Generated 2026-09-22T17:28Z. Schema 1.2.1. Snapshots so far: 2026-09-18, 2026-09-19, 2026-09-20, 2026-09-21, 2026-09-22.

Tables

tablerowscoverswhat
market_daily932026-06-21 to 2026-09-21OpenRouter, one row per complete UTC day: tokens served, estimated dollars, concentration (HHI, top 3) and the attested share, by tokens and by estimated dollars.
provider_daily1,4022026-06-21 to 2026-09-21OpenRouter, one row per day and provider: share of tokens and of estimated dollars.
provider_snapshot4402026-09-18 to 2026-09-22OpenRouter providers as listed on each collection day: tokens a day and a month, models, attested flag, and TrustedRouter's measurements of the same provider (tr_*).
model_provider_window2,9152026-09-20 to 2026-09-22OpenRouter, tokens a day for each model on each provider over 1, 7 and 30 complete days (the top 60 models of each window).
openrouter_endpoint_quote1,1722026-09-18 to 2026-09-22Listed prices per endpoint of OpenRouter's top 20 models on each collection day: prompt, completion, cache read, discount, quantization.
openrouter_model_week1002026-09-18 to 2026-09-22OpenRouter's weekly top 20 models: tokens, requests, share, open weights.
openrouter_app_week1002026-09-18 to 2026-09-22OpenRouter's weekly top 20 apps: tokens and requests.
vercel_share_daily2,1832026-07-20 to 2026-09-22Vercel AI Gateway, daily share of tokens by model lab.
vercel_leaderboard_daily14,2192026-07-21 to 2026-09-22Vercel AI Gateway's leaderboard export in full: daily share per model and per lab, for tokens, requests, spend, images and videos.
vercel_endpoint1,7082026-09-20 to 2026-09-22Vercel AI Gateway, every language model on every provider on each collection day: listed prices, uptime, latency, throughput.
vercel_rank802026-09-20 to 2026-09-22Vercel AI Gateway's two top-10 rankings of apps and providers, by tokens and by spend.
price_compare3332026-09-20 to 2026-09-22The same model on the same provider, quoted by OpenRouter and by Vercel: listed input price per million tokens.
trustedrouter_provider3372026-09-18 to 2026-09-22TrustedRouter's own measurements of providers on each collection day: first-token time, availability, completion, samples.
attested_registry852026-09-18 to 2026-09-22The awesome-private-inference registry's attestation checks per endpoint: status and layers proven.
chutes_daily162026-09-07 to 2026-09-22Chutes, by its own meter: dollars metered a day and the share on TEE chutes.
openrouter_endpoint_stats6,2142026-09-19 to 2026-09-22Every endpoint of every OpenRouter catalog model (about 450 models, 1,500 endpoints) on each collection day: listed prices as in openrouter_endpoint_quote, plus OpenRouter's one-month effective prices, cache-hit rate and tokens per endpoint.
openrouter_endpoint_daily18,7782026-09-12 to 2026-09-22OpenRouter's daily figures per endpoint for every catalog model: cache-hit rate, median throughput, median latency, uptime.
openrouter_model_daily12,5152026-08-20 to 2026-09-22OpenRouter's daily activity per catalog model and variant: prompt, completion, cached and reasoning tokens, requests, tool calls.
openrouter_model_app7,0232026-09-19 to 2026-09-22OpenRouter's top apps for each catalog model on each collection day (up to five per model): the app by model matrix.
probe_request11,7992026-09-19 to 2026-09-22MEASURED BY US.
quote_change1,245MEASURED BY US.
board_entry2,852,6932023-02-24 to 2026-09-30Every leaderboard we read, in one long table: one row per board, period, entity and metric, for each value it has had.
source40Every leaderboard source we read, public or not: who runs it, the provenance tier of its rows, its license, the citation it asks for, whether its rows may be redistributed, how and how often we read it.
board412Every board of every source, public or not: what it ranks, its period kind, depth, what it measures in one sentence, and the license that applies to it.
board_metric833One row per board and metric: the unit of board_entry.value for that metric and which direction is better.
board_coverage405One row per board, public or not: rows held, rows whose entity is mapped to a canonical id, first and last period, first and last snapshot, last collection time.
name_map26,157How every published name was resolved: one row per source, entity type, raw name, provider and join key, with the canonical ids boards/names.py gives it.
app162Canonical apps: OpenRouter's app slug where it has one, else the title as a slug.
app_alias172Every spelling of an app that a source uses.
family206Model families (meta/llama-3.1): the target of family aliases, for boards that name a family and not one model (Ollama pulls a library entry, not a checkpoint).
family_alias206Every spelling of a family that a source uses, from data/curated/aliases/family/.
gateway27Canonical gateways and routers that sit in front of providers (OpenRouter, Vercel AI Gateway, Cloudflare AI Gateway): the entities of gateway benchmark boards such as ComputeSDK.
gateway_alias15Every spelling of a gateway that a source uses, from data/curated/aliases/gateway/.
hardware21Canonical accelerators and systems (nvidia-h200, nvidia-b300): the entities of model by hardware boards (MLPerf) and the target of board_entry.hardware_id.
hardware_alias31Every spelling of a hardware that a source uses, from data/curated/aliases/hardware/.
lab430Canonical labs: the part of the OpenRouter slug before the slash.
lab_alias87Every spelling of a lab that a source uses.
model1,896Canonical models.
model_alias9,714Every spelling of a model that a source uses, and the canonical model it maps to.
provider188Canonical inference providers: OpenRouter's provider slug.
provider_alias700Every spelling of a provider that a source uses.
provider_model_alias9,264A provider's native model id mapped to the canonical model: the crosswalk that resolves price and speed boards, where a row is a model on a provider.
task29Canonical task categories (agent:memory_extraction, code-debugging): the entities of task boards such as OpenRouter's task classifications.

market_daily

OpenRouter, one row per complete UTC day: tokens served, estimated dollars, concentration (HHI, top 3) and the attested share, by tokens and by estimated dollars.

93 rows, date 2026-06-21 to 2026-09-21. Key: date, router.

columntypeunittierdescription
datedatedatereportedUTC day the row is about. The collection day itself is partial and left out. [key]
routervarcharidreportedAlways openrouter. Nothing is summed across routers. [key]
tokensdoubletokensreportedTokens served that day, summed over every provider's 90-day token chart.
usd_estdoubleusdderivedEstimated dollars: tokens times the blended one-month effective price of each model at each provider.
hhi_tokensdoubleindexderivedHerfindahl index of provider shares of tokens, 0 to 10,000.
hhi_usd_estdoubleindexderivedHerfindahl index of provider shares of estimated dollars, 0 to 10,000.
top3_tokensdoublefractionderivedShare of tokens served by the three largest providers.
top3_usd_estdoublefractionderivedShare of estimated dollars earned by the three largest providers.
attested_share_tokensdoublefractionderivedShare of tokens served by attested providers.
attested_share_usd_estdoublefractionderivedShare of estimated dollars earned by attested providers.
usd_priced_token_sharedoublefractionderivedShare of the day's tokens that had a model attached and so a price of their own; the rest is priced at the provider's mean.

provider_daily

OpenRouter, one row per day and provider: share of tokens and of estimated dollars. Covers the largest providers and the attested ones; the rest are inside the daily total only.

1,402 rows, date 2026-06-21 to 2026-09-21. Key: date, router, provider.

columntypeunittierdescription
datedatedatereportedUTC day the row is about. [key]
routervarcharidreportedAlways openrouter. [key]
providervarcharidreportedProvider, as the OpenRouter provider slug (the one spelling every table uses). [key; joins provider_snapshot.provider]
share_tokensdoublefractionderivedProvider's share of the day's tokens.
share_usd_estdoublefractionderivedProvider's share of the day's estimated dollars.

provider_snapshot

OpenRouter providers as listed on each collection day: tokens a day and a month, models, attested flag, and TrustedRouter's measurements of the same provider (tr_*). The closest thing to a provider dimension until the entity tables exist.

440 rows, snapshot 2026-09-18 to 2026-09-22. Key: snapshot, router, provider.

columntypeunittierdescription
snapshotdatedatereportedUTC day we collected the row. [key]
routervarcharidreportedAlways openrouter. [key]
providervarcharidreportedProvider, as the OpenRouter provider slug (the one spelling every table uses). [key]
provider_namevarchartextreportedProvider display name, as OpenRouter writes it. [not null]
namevarchartextreportedDeprecated, use provider_name. Provider display name.
attestedbooleanflagderivedTrue when the provider is one of the attested (TEE) providers: Phala, Venice, Chutes, NEAR AI, Tinfoil, RedPill.
modelsbigintcountreportedModels the provider serves on OpenRouter.
tokens_daybiginttokensreportedTokens over the rolling day, from the providers page.
tokens_30dbiginttokensreportedTokens over the rolling 30 days, from the providers page.
share_daydoublefractionderivedProvider's share of all providers' rolling-day tokens.
tr_ttft_p50_msdoublemsmeasured by othersTrustedRouter's median time to first token for this provider.
tr_availabilitydoublefractionmeasured by othersTrustedRouter's availability for this provider.
tr_samplesbigintcountmeasured by othersRequests behind TrustedRouter's figures.

model_provider_window

OpenRouter, tokens a day for each model on each provider over 1, 7 and 30 complete days (the top 60 models of each window).

2,915 rows, snapshot 2026-09-20 to 2026-09-22. Key: snapshot, router, period_days, model, provider.

columntypeunittierdescription
snapshotdatedatereportedUTC day we collected the row. [key]
routervarcharidreportedAlways openrouter. [key]
period_daysbigintdaysreportedLength of the window: 1, 7 or 30 complete UTC days. [key]
period_startdatedatereportedFirst UTC day of the window, inclusive.
period_enddatedatereportedLast UTC day of the window, inclusive.
windowvarchartextreportedDeprecated, use period_days. The window as 1d, 7d or 30d. A reserved word in SQL.
firstdatedatereportedDeprecated, use period_start. First day of the window.
lastdatedatereportedDeprecated, use period_end. Last day of the window.
modelvarcharidreportedModel, as the OpenRouter model id (the one spelling every table uses). [key]
labvarchartextderivedLab that built the model, display name, normalized by our lab map.
providervarcharidreportedProvider, as the OpenRouter provider slug (the one spelling every table uses). [key; joins provider_snapshot.provider]
provider_namevarchartextreportedProvider display name.
attestedbooleanflagderivedTrue when the provider is one of the attested (TEE) providers: Phala, Venice, Chutes, NEAR AI, Tinfoil, RedPill.
tokens_daydoubletokens_per_dayderivedTokens a day of this model on this provider, averaged over the window.
share_of_modeldoublefractionderivedProvider's share of the model's tokens over the window.

openrouter_endpoint_quote

Listed prices per endpoint of OpenRouter's top 20 models on each collection day: prompt, completion, cache read, discount, quantization. Paid endpoints of the standard variant, read from the model pages. For every catalog model see openrouter_endpoint_stats.

1,172 rows, snapshot 2026-09-18 to 2026-09-22. Key: snapshot, model, endpoint.

columntypeunittierdescription
snapshotdatedatereportedUTC day we collected the row. [key]
modelvarcharidreportedModel, as the OpenRouter model id (the one spelling every table uses). [key]
labvarchartextderivedLab that built the model, display name, normalized by our lab map.
providervarcharidreportedProvider, as the OpenRouter provider slug (the one spelling every table uses). Through schema 1.0 this column held the display name, which is now provider_name. [joins provider_snapshot.provider]
provider_namevarchartextreportedProvider display name, as OpenRouter writes it.
endpointvarcharidreportedOpenRouter endpoint tag: the provider slug, then a quantization, region or service tier where the provider has several (deepinfra/fp8). [key]
prompt_usd_per_mdoubleusd_per_mreportedListed price of a million prompt tokens.
completion_usd_per_mdoubleusd_per_mreportedListed price of a million completion tokens.
cache_read_usd_per_mdoubleusd_per_mreportedListed price of a million cached prompt tokens; null where the endpoint lists none.
discountdoublefractionreportedDiscount against list that the provider declares.
quantizationvarchartextreportedQuantization the endpoint declares (fp8, fp4, bf16, unknown).
derankedbooleanflagreportedOpenRouter's flag for an endpoint it routes to last.
attestedbooleanflagderivedTrue when the provider is one of the attested (TEE) providers: Phala, Venice, Chutes, NEAR AI, Tinfoil, RedPill.

openrouter_model_week

OpenRouter's weekly top 20 models: tokens, requests, share, open weights.

100 rows, snapshot 2026-09-18 to 2026-09-22. Key: snapshot, model.

columntypeunittierdescription
snapshotdatedatereportedUTC day we collected the row. [key]
period_startdatedatereportedFirst UTC day of the week, inclusive.
period_enddatedatereportedLast UTC day of the week, inclusive.
windowvarchartextreportedDeprecated, use period_end. The week as prose (week ending 2026-09-20). A reserved word in SQL.
modelvarcharidreportedModel, as the OpenRouter model id (the one spelling every table uses). [key]
labvarchartextderivedLab that built the model, display name, normalized by our lab map.
open_weightsbooleanflagderivedTrue when the catalog links a Hugging Face repository for the model.
tokens_weekbiginttokensreportedPrompt plus completion tokens over the week.
requests_weekbigintrequestsreportedRequests over the week.
sharedoublefractionderivedModel's share of the top 20's tokens.

openrouter_app_week

OpenRouter's weekly top 20 apps: tokens and requests.

100 rows, snapshot 2026-09-18 to 2026-09-22. Key: snapshot, app.

columntypeunittierdescription
snapshotdatedatereportedUTC day we collected the row. [key]
period_startdatedatereportedFirst UTC day of the week, inclusive.
period_enddatedatereportedLast UTC day of the week, inclusive.
windowvarchartextreportedDeprecated, use period_end. The week as prose. A reserved word in SQL.
appvarchartextreportedApp title, as OpenRouter writes it. [key]
categoriesvarchar[]listreportedOpenRouter's category tags for the app. Through schema 1.0 a comma-joined string.
tokens_weekbiginttokensreportedTokens over the week.
requests_weekbigintrequestsreportedRequests over the week.

vercel_share_daily

Vercel AI Gateway, daily share of tokens by model lab. Shares only: Vercel publishes no totals. A subset of vercel_leaderboard_daily (lab, tokens) with lab display names, kept as it was for existing queries.

2,183 rows, date 2026-07-20 to 2026-09-22. Key: date, router, level, metric, name.

columntypeunittierdescription
datedatedatereportedUTC day the share is about. [key]
routervarcharidreportedAlways vercel. [key]
levelvarcharidreportedAlways lab. [key]
metricvarcharidreportedAlways tokens. [key]
namevarchartextreportedLab that built the model, display name, normalized by our lab map. [key]
sharedoublefractionreportedLab's share of the gateway's tokens that day.
snapshotdatedatereportedUTC day of the collection the value was read from (the newest that holds the date). snapshot = date means the day was still running. [not null]

vercel_leaderboard_daily

Vercel AI Gateway's leaderboard export in full: daily share per model and per lab, for tokens, requests, spend, images and videos. Upstream keeps a rolling 61 days; this table keeps every day ever collected. Models are the ones Vercel names (about ten a day per metric); the rest are inside the row named Other.

14,219 rows, date 2026-07-21 to 2026-09-22. Key: date, level, metric, name. Keeps every day ever collected.

columntypeunittierdescription
datedatedatereportedUTC day the share is about. [key]
routervarcharidreportedAlways vercel.
levelvarcharidreportedmodel or lab. [key]
metricvarcharidreportedtokens, requests, spend, images or videos (Vercel writes imageCount and videoCount; see metric_raw). [key]
metric_rawvarcharidreportedMetric as Vercel writes it. [not null]
namevarchartextreportedModel display name or lab slug, exactly as Vercel writes it (GLM 5.3 Flash; zai). [key]
labvarchartextderivedFor lab rows, the lab display name by our lab map (zai becomes Z.ai). Null for model rows: mapping Vercel's model names waits for the alias tables.
residualbooleanflagderivedTrue for the row named Other: everything Vercel did not name in that export. The named set moves as the 61-day window rolls, so after the union the named rows of a day can add up to more than 100 minus Other; 1 - sum(share of the named rows) is always exact. [not null]
sharedoublefractionreportedShare of the gateway's total for that metric and day (share_percent / 100). [not null]
snapshotdatedatereportedUTC day of the collection this value was read from: the newest one that published it. A value with snapshot = date was read while that day was still running and is partial; it is replaced the next night.

vercel_endpoint

Vercel AI Gateway, every language model on every provider on each collection day: listed prices, uptime, latency, throughput. Through schema 1.0 it held the newest day only; it now keeps every day from 2026-09-20 on.

1,708 rows, snapshot 2026-09-20 to 2026-09-22. Key: snapshot, model_raw, provider_raw. Keeps every day ever collected.

columntypeunittierdescription
snapshotdatedatereportedUTC day we collected the row. [key]
routervarcharidreportedAlways vercel.
modelvarcharidderivedModel, as the OpenRouter model id (the one spelling every table uses). Only the lab prefix zai/ is rewritten (to z-ai/); other Vercel ids are as published until the alias tables exist.
model_rawvarcharidreportedModel id exactly as Vercel writes it (zai/glm-5.3). [key]
providervarcharidderivedProvider, as the OpenRouter provider slug (the one spelling every table uses). Vercel's bedrock, vertex, google, zai, togetherai and moonshot are mapped; the rest are as published. [joins provider_snapshot.provider]
provider_rawvarcharidreportedProvider exactly as Vercel writes it (bedrock). [key]
prompt_usd_per_mdoubleusd_per_mreportedListed price of a million prompt tokens.
completion_usd_per_mdoubleusd_per_mreportedListed price of a million completion tokens.
cache_read_usd_per_mdoubleusd_per_mreportedListed price of a million cached prompt tokens.
quantizationvarchartextreportedQuantization, where Vercel states one (so far it never does).
zdrbooleanflagreportedVercel lists the endpoint as zero data retention.
implicit_cachingbooleanflagreportedVercel lists the endpoint as caching prompts without being asked.
uptime_1ddoublepctreportedUptime over the last day, 0 to 100, as Vercel reports it.
latency_p50_msdoublemsreportedMedian latency over the last hour, as Vercel reports it.
throughput_p50doubletokens_per_sreportedMedian throughput over the last hour, as Vercel reports it.

vercel_rank

Vercel AI Gateway's two top-10 rankings of apps and providers, by tokens and by spend. Ranks only.

80 rows, snapshot 2026-09-20 to 2026-09-22. Key: snapshot, kind, name.

columntypeunittierdescription
snapshotdatedatereportedUTC day we collected the row. [key]
kindvarcharidreportedapp or provider. [key]
namevarchartextreportedApp or provider display name, as Vercel writes it. [key]
rank_by_tokensbigintrankreportedRank by token volume; null when outside that top 10.
rank_by_spendbigintrankreportedRank by spend; null when outside that top 10.

price_compare

The same model on the same provider, quoted by OpenRouter and by Vercel: listed input price per million tokens.

333 rows, snapshot 2026-09-20 to 2026-09-22. Key: snapshot, model, provider.

columntypeunittierdescription
snapshotdatedatederivedUTC day we collected the row. [key]
modelvarcharidderivedModel, as the OpenRouter model id (the one spelling every table uses). [key]
providervarcharidderivedProvider, as the OpenRouter provider slug (the one spelling every table uses). [key; joins provider_snapshot.provider]
provider_namevarchartextderivedProvider display name, as OpenRouter writes it.
openrouter_endpointvarcharidderivedThe OpenRouter endpoint compared: the provider's plain one where it has one, else its cheapest variant. [joins openrouter_endpoint_quote.endpoint]
openrouter_prompt_usd_per_mdoubleusd_per_mreportedOpenRouter's listed prompt price.
vercel_prompt_usd_per_mdoubleusd_per_mreportedVercel's listed prompt price.
vercel_vs_openrouterdoublefractionderivedVercel's price over OpenRouter's, minus one.
openrouter_cache_usd_per_mdoubleusd_per_mreportedOpenRouter's listed cache-read price.
vercel_cache_usd_per_mdoubleusd_per_mreportedVercel's listed cache-read price.

trustedrouter_provider

TrustedRouter's own measurements of providers on each collection day: first-token time, availability, completion, samples. Measured by TrustedRouter, not by us.

337 rows, snapshot 2026-09-18 to 2026-09-22. Key: snapshot, provider_raw.

columntypeunittierdescription
snapshotdatedatemeasured by othersUTC day we collected the row. [key]
providervarcharidderivedProvider, as the OpenRouter provider slug (the one spelling every table uses). Through schema 1.0 this column held TrustedRouter's own slug, which is now provider_raw; the two differ for the nine providers of the TR_SLUG map in build_dataset.py (zai, gmi, kimi, grok, nvidia-nim and others). [joins provider_snapshot.provider]
provider_rawvarcharidmeasured by othersProvider slug exactly as TrustedRouter writes it. [key]
rankbigintrankmeasured by othersTrustedRouter's rank; null for providers it does not qualify.
ttft_p50_msdoublemsmeasured by othersMedian time to first token.
ttft_p95_msdoublemsmeasured by others95th percentile time to first token.
throughput_tpsdoubletokens_per_smeasured by othersThroughput, as TrustedRouter measures it.
completiondoublefractionmeasured by othersShare of requests that completed.
availabilitydoublefractionmeasured by othersShare of requests the provider accepted.
samplesbigintcountmeasured by othersRequests behind the figures.
modelsbigintcountmeasured by othersModels measured.
verified_confidentialbooleanflagmeasured by othersTrustedRouter labels the provider as verified confidential.
zdr_labelbooleanflagmeasured by othersTrustedRouter labels the provider as zero data retention.

attested_registry

The awesome-private-inference registry's attestation checks per endpoint: status and layers proven.

85 rows, snapshot 2026-09-18 to 2026-09-22. Key: snapshot, provider, endpoint.

columntypeunittierdescription
snapshotdatedatemeasured by othersUTC day we collected the row. [key]
providervarcharidmeasured by othersProvider, as the registry writes it; its ids match OpenRouter's slugs where the provider is on OpenRouter. [key]
endpointvarcharidmeasured by othersThe model or endpoint name the registry checks. [key]
statusvarchartextmeasured by othersRegistry verdict: verified, partial, invalid.
layers_provenbigintcountmeasured by othersAttestation layers the registry could prove.
layers_requiredbigintcountmeasured by othersLayers it requires.
missingvarchartextmeasured by othersLayers not proven, comma-joined.

chutes_daily

Chutes, by its own meter: dollars metered a day and the share on TEE chutes.

16 rows, date 2026-09-07 to 2026-09-22. Key: date.

columntypeunittierdescription
datedatedatereportedUTC day. The newest day is partial. [key]
usd_metereddoubleusdreportedDollars Chutes metered that day.
share_teedoublefractionderivedShare of those dollars on chutes whose name marks them as TEE.
share_plaindoublefractionderivedShare on chutes listed without TEE.
share_unlisteddoublefractionderivedShare on chutes missing from the public list.

openrouter_endpoint_stats

Every endpoint of every OpenRouter catalog model (about 450 models, 1,500 endpoints) on each collection day: listed prices as in openrouter_endpoint_quote, plus OpenRouter's one-month effective prices, cache-hit rate and tokens per endpoint. From the frontend stats behind the model pages. Model ids that start with ~ are OpenRouter's "latest" aliases and repeat another model's endpoints.

6,214 rows, snapshot 2026-09-19 to 2026-09-22. Key: snapshot, model_permaslug, variant, endpoint_id. Keeps every day ever collected.

columntypeunittierdescription
snapshotdatedatereportedUTC day we collected the row. [key]
modelvarcharidreportedModel, as the OpenRouter model id (the one spelling every table uses). With the variant as a suffix where it is not standard (:free, :batch). [not null]
model_permaslugvarcharidreportedOpenRouter's dated permanent slug for the model. [key]
variantvarcharidreportedstandard, free or batch. [key]
labvarchartextderivedLab that built the model, display name, normalized by our lab map.
providervarcharidreportedProvider, as the OpenRouter provider slug (the one spelling every table uses). [joins provider_snapshot.provider]
provider_namevarchartextreportedProvider display name.
endpointvarcharidreportedOpenRouter endpoint tag (deepinfra/fp8).
endpoint_idvarcharidreportedOpenRouter's id for the endpoint; the key of openrouter_endpoint_daily. [key]
prompt_usd_per_mdoubleusd_per_mreportedListed price of a million prompt tokens.
completion_usd_per_mdoubleusd_per_mreportedListed price of a million completion tokens.
cache_read_usd_per_mdoubleusd_per_mreportedListed price of a million cached prompt tokens.
discountdoublefractionreportedDiscount against list that the provider declares.
quantizationvarchartextreportedQuantization the endpoint declares.
derankedbooleanflagreportedOpenRouter's flag for an endpoint it routes to last.
attestedbooleanflagderivedTrue when the provider is one of the attested (TEE) providers: Phala, Venice, Chutes, NEAR AI, Tinfoil, RedPill.
is_freebooleanflagreportedA free endpoint.
is_disabledbooleanflagreportedOpenRouter lists the endpoint as disabled.
top20booleanflagderivedThe model was in OpenRouter's weekly top 20 that day. [not null]
effective_prompt_usd_per_mdoubleusd_per_mreportedOpenRouter's effective price of a million prompt tokens over the last month, after cache hits and discounts.
effective_completion_usd_per_mdoubleusd_per_mreportedEffective price of a million completion tokens over the last month.
cache_hit_rate_1mdoublefractionreportedCache-hit rate over the last month, as OpenRouter reports it.
tokens_1mbiginttokensreportedTokens the endpoint served over the last month.

openrouter_endpoint_daily

OpenRouter's daily figures per endpoint for every catalog model: cache-hit rate, median throughput, median latency, uptime. Upstream keeps about 8 days (3 for uptime); this table keeps every day ever collected.

18,778 rows, date 2026-09-12 to 2026-09-22. Key: date, model_permaslug, variant, endpoint_id. Keeps every day ever collected.

columntypeunittierdescription
datedatedatereportedUTC day the figures are about. [key]
modelvarcharidreportedModel, as the OpenRouter model id (the one spelling every table uses). [not null]
model_permaslugvarcharidreportedOpenRouter's dated permanent slug. [key]
variantvarcharidreportedstandard, free or batch. [key]
endpoint_idvarcharidreportedOpenRouter's id for the endpoint. The literal openrouter marks OpenRouter's own uptime for the model across all its endpoints. [key; joins openrouter_endpoint_stats.endpoint_id]
providervarcharidreportedProvider, as the OpenRouter provider slug (the one spelling every table uses). Null on the endpoint_id = openrouter rows, and for an endpoint that had left every endpoint list by the time we collected. [joins provider_snapshot.provider]
endpointvarcharidreportedOpenRouter endpoint tag (deepinfra/fp8).
cache_hit_ratedoublefractionreportedShare of prompt tokens served from cache that day, as OpenRouter reports it (published in percent).
throughput_p50_tpsdoubletokens_per_sreportedMedian throughput that day.
latency_p50_msdoublemsreportedMedian latency to first token that day.
uptimedoublefractionreportedUptime that day (published in percent).
snapshotdatedatereportedUTC day of the collection this value was read from: the newest one that published it. A value with snapshot = date was read while that day was still running and is partial; it is replaced the next night.

openrouter_model_daily

OpenRouter's daily activity per catalog model and variant: prompt, completion, cached and reasoning tokens, requests, tool calls. Upstream keeps 31 days; this table keeps every day ever collected.

12,515 rows, date 2026-08-20 to 2026-09-22. Key: date, model_permaslug, variant. Keeps every day ever collected.

columntypeunittierdescription
datedatedatereportedUTC day the figures are about. [key]
modelvarcharidreportedModel, as the OpenRouter model id (the one spelling every table uses). [not null]
model_permaslugvarcharidreportedOpenRouter's dated permanent slug. [key]
variantvarcharidreportedstandard, free or batch. [key]
labvarchartextderivedLab that built the model, display name, normalized by our lab map.
prompt_tokensbiginttokensreportedPrompt tokens that day.
completion_tokensbiginttokensreportedCompletion tokens that day.
cached_tokensbiginttokensreportedPrompt tokens served from cache (native count).
reasoning_tokensbiginttokensreportedReasoning tokens (native count).
requestsbigintrequestsreportedRequests that day.
tool_callsbigintcountreportedTool calls that day.
tool_call_error_requestsbigintrequestsreportedRequests with a tool-call error.
media_promptsbigintcountreportedImages and other media sent in prompts.
snapshotdatedatereportedUTC day of the collection this value was read from: the newest one that published it. A value with snapshot = date was read while that day was still running and is partial; it is replaced the next night.

openrouter_model_app

OpenRouter's top apps for each catalog model on each collection day (up to five per model): the app by model matrix. The window is undocumented upstream and longer than a week.

7,023 rows, snapshot 2026-09-19 to 2026-09-22. Key: snapshot, model_permaslug, variant, app_id. Keeps every day ever collected.

columntypeunittierdescription
snapshotdatedatereportedUTC day we collected the row. [key]
modelvarcharidreportedModel, as the OpenRouter model id (the one spelling every table uses). [not null]
model_permaslugvarcharidreportedOpenRouter's dated permanent slug. [key]
variantvarcharidreportedstandard, free or batch. [key]
app_idbigintidreportedOpenRouter's id for the app. [key]
appvarchartextreportedApp title, as OpenRouter writes it.
categoriesvarchar[]listreportedOpenRouter's category tags for the app.
tokensbiginttokensreportedTokens of this model used by the app over OpenRouter's window.
requestsbigintrequestsreportedRequests over the same window.

probe_request

MEASURED BY US. One row per request the prober sent, through OpenRouter, TrustedRouter or Vercel (route): tokens, cached tokens, the bill, first-token time, who served it. No prompt or response text. A cache pair is two rows with the same session_id, repeat 1 and 2. Early and unreviewed.

11,799 rows, date 2026-09-19 to 2026-09-22. Key: ts, session_id.

columntypeunittierdescription
tstimestamptimestampmeasured by usWhen the request was sent, UTC. [key]
datedatedatemeasured by usUTC day of ts.
routevarcharidmeasured by usRouter the request went through: openrouter, trustedrouter or vercel.
jobvarcharidmeasured by usProber job that sent it.
run_idvarcharidmeasured by usOne run of a job.
session_idvarcharidmeasured by usOne conversation; the rows of a cache pair or of an agent trace share it. [key]
workloadvarcharidmeasured by usWorkload: cache-pair, agent-trace and the rest of the prober's taxonomy.
familyvarcharidmeasured by usWorkload family.
modelvarcharidmeasured by usModel, as the OpenRouter model id (the one spelling every table uses).
pinnedbooleanflagmeasured by usTrue when the request named one provider and allowed no fallback.
provider_requestedvarcharidmeasured by usWhat the prober asked the router for, in that router's own spelling: an OpenRouter endpoint tag (deepinfra/fp8), or a provider slug on the other routes. Null when unpinned.
providervarcharidderivedProvider, as the OpenRouter provider slug (the one spelling every table uses). From provider_requested; null when unpinned. [joins provider_snapshot.provider]
endpointvarcharidderivedOpenRouter endpoint tag requested (deepinfra/fp8); null on other routes and when unpinned. [joins openrouter_endpoint_stats.endpoint]
provider_servedvarchartextreportedProvider the router says served the request, display name as the router writes it.
turnbigintcountmeasured by usTurn within a multi-turn session.
repeatbigintcountmeasured by us1 for the first send, 2 for the same bytes again.
armvarcharidmeasured by usExperiment arm.
request_namevarcharidmeasured by usName of the request within its workload.
event_idvarcharidmeasured by usEvent the request belongs to.
repeat_gap_sdoublesmeasured by usSeconds between the first send and the repeat.
repeated_prompt_tokensbiginttokensmeasured by usPrompt tokens that repeat an earlier request in the session.
prompt_tokensbiginttokensreportedPrompt tokens, as the router's usage block reports.
completion_tokensbiginttokensreportedCompletion tokens, as reported.
cached_tokensbiginttokensreportedPrompt tokens the provider recognized from cache, as reported.
reasoning_tokensbiginttokensreportedReasoning tokens, as reported.
billed_usddoubleusdmeasured by usWhat the request cost.
bill_reported_by_routerbooleanflagmeasured by usTrue when the router reported the bill; false when we computed it from tokens and quotes.
quote_prompt_usd_per_mdoubleusd_per_mreportedListed prompt price at send time.
quote_cache_read_usd_per_mdoubleusd_per_mreportedListed cache-read price at send time.
quantizationvarchartextreportedQuantization the endpoint declared at send time.
ttft_msdoublemsmeasured by usTime to first token, measured at the prober.
total_msdoublemsmeasured by usTime to the last byte.
tokens_per_sdoubletokens_per_smeasured by usCompletion tokens per second after the first token.
finish_reasonvarchartextreportedFinish reason the router returned.
http_statusbigintidmeasured by usHTTP status.
failedbooleanflagmeasured by usThe request errored.
check_passedbooleanflagmeasured by usThe answer passed the workload's check; null where there is none.
workload_hashvarcharidmeasured by usHash of the workload file the request was built from: equal hashes mean the same workload version.
response_sha256varcharidmeasured by usSHA-256 of the response text. The text itself is never exported.

quote_change

MEASURED BY US. Every change in a quoted price that the 5-minute poll of OpenRouter saw, for the panel models.

1,245 rows. Key: ts, model, endpoint, field.

columntypeunittierdescription
tstimestamptimestampmeasured by usWhen the poll saw the change, UTC. [key]
modelvarcharidmeasured by usModel, as the OpenRouter model id (the one spelling every table uses). [key]
providervarcharidderivedProvider, as the OpenRouter provider slug (the one spelling every table uses). [not null; joins provider_snapshot.provider]
endpointvarcharidmeasured by usOpenRouter endpoint tag (deepinfra/fp8). [key; joins openrouter_endpoint_stats.endpoint]
fieldvarcharidmeasured by usWhat changed: prompt_usd_per_m, completion_usd_per_m, cache_read_usd_per_m, or endpoint (listed or delisted). [key]
olddoubleusd_per_mmeasured by usValue before, for price fields; null for field = endpoint. Through schema 1.0 typed JSON.
newdoubleusd_per_mmeasured by usValue after, for price fields; null for field = endpoint. Through schema 1.0 typed JSON.
old_textvarchartextmeasured by usValue before, as text, for every field (null, listed).
new_textvarchartextmeasured by usValue after, as text, for every field (listed, delisted).

board_entry

Every leaderboard we read, in one long table: one row per board, period, entity and metric, for each value it has had. Usage shares, arena ratings, benchmark scores, prices and measured speeds from about 40 sources, with history where upstream keeps it. One Parquet file per source and year (board_entry/source=<id>/year=<yyyy>/part.parquet, listed in board_entry_files.json). A source publishes here only if its license allows redistribution; the rest is collected into files that are not served. The tier of a row is its source's (source.tier).

2,852,693 rows, period_end 2023-02-24 to 2026-09-30. Key: board_id, period_end, entity_raw, provider_raw, effort, agent, variant, hardware, metric, snapshot. Keeps every day ever collected. 32 files over 12 sources, listed in board_entry_files.json; snapshots 2024-10-22 to 2026-09-22; the private copy holds 3,718,227 rows of 38 sources.

columntypeunittierdescription
source_idvarcharidper sourceSource the row came from: the first segment of board_id. Also the directory the file lives in (source=<source_id>). [not null; joins source.id]
board_idvarcharidper sourceBoard the row belongs to, <source_id>.<name>. Part of the row key. [key; joins board.id]
snapshotdatedateper sourceUTC day we first collected this value for this key: the snapshot that is kept. A key gets a new row only when its value, interval, rank, n or cost changes, so most sources, which publish the same history again every night, cost one row per value and not one per night. Part of the row key. [key]
first_seendatedateper sourceThe same day as snapshot, under the name that pairs with last_seen. [not null]
last_seendatedateper sourceNewest UTC day on which we collected the same value for this key. Earlier than the newest snapshot of the board when the entry left the board or its value changed (the next row of the key says which). [not null]
collected_attimestamptimestampper sourceUTC time of the collection on first_seen. [not null]
last_collected_attimestamptimestampper sourceUTC time of the collection on last_seen. [not null]
period_startdatedateper sourceFirst day the value covers; null on boards dated as of a day (asof) and on boards with no date of their own.
period_enddatedateper sourceLast day the value covers, or the day upstream published it; null on boards that only show "now", which are told apart by snapshot. The file's year= directory is the year of period_end, else of snapshot. Part of the row key. [key]
rankintegerrankper sourcePosition on the board for this metric and period, 1 is top; null for residual rows and unranked values.
entity_typevarcharidper sourceWhat is ranked: model, lab, app, provider, endpoint, router, gateway, hardware, task, language, country, system, or residual (the row named Other, kept so shares still sum to 100). [not null]
entity_rawvarchartextper sourceName exactly as the source published it, never edited. Part of the row key. [key]
entity_idvarcharidderivedCanonical id of the entity: joins the table named by entity_type (model.id, lab.id, provider.id, app.id, gateway.id, task.id, hardware.id). A model row that no model alias answers may resolve to a family (family.id, name_map.method = family). Null until the name is mapped. Resolved when the tables are built, from data/curated/aliases/, so a corrected alias applies to all history. [joins model.id]
provider_rawvarchartextper sourceServing provider exactly as published, on boards whose row is a model on a provider (prices, speeds, uptime). Part of the row key. [key]
provider_idvarcharidderivedCanonical provider id; null until mapped. [joins provider.id]
effortvarchartextper sourceThinking mode or reasoning effort as published (high, thinking, 16k). Part of the row key. [key]
agentvarchartextper sourceAgent scaffold or harness the model ran in (SWE-agent, Terminus 2). Part of the row key. [key]
variantvarchartextper sourceAnything else that splits rows of one entity: free, batch, quantization, checkpoint date, scenario. Part of the row key. [key]
variant_from_namebooleanflagderivedTrue when the source published no variant and the name mapping took one off the name (:free, a checkpoint date). [not null]
hardwarevarchartextper sourceAccelerator or system, on model by hardware boards. Part of the row key. [key]
hardware_idvarcharidderivedCanonical id of the hardware column (8x NVIDIA H200-SXM-141GB gives nvidia-h200; a system that mixes accelerators has none). Not in the row contract: derived at build time through boards/names.py, like entity_id. [joins hardware.id]
metricvarcharidper sourceMetric name, declared per board in board_metric. Part of the row key. [key; joins board_metric.metric]
valuedoubleper_rowper sourceThe number, in the unit of the unit column. Percentages are 0 to 100; prices are dollars per million tokens. [not null]
value_lowdoubleper_rowper sourceLower bound of the published interval.
value_highdoubleper_rowper sourceUpper bound of the published interval.
unitvarcharidper sourceUnit of value: tokens, requests, usd, usd_per_m, share_pct, rank, rating, pct, index, runs, downloads, ms, s, minutes, tokens_per_s or count. [not null]
nbigintcountper sourceVotes, samples or trials behind the value.
cost_usddoubleusdper sourceCost per task or run in dollars, where the board reports it next to the score.
source_urlvarcharurlper sourceURL the raw bytes came from, on first_seen. Never holds a key. [not null]
extrajsonjsonper sourceSmall flat object of leftovers the source published (organization, license, upstream ids). partial = true marks a period that was still running when collected.

source

Every leaderboard source we read, public or not: who runs it, the provenance tier of its rows, its license, the citation it asks for, whether its rows may be redistributed, how and how often we read it.

40 rows. Key: id. Also as source.csv.gz. Every source and board, public or not.

columntypeunittierdescription
idvarcharidderivedSource id, the first segment of every board_id. [key]
namevarchartextderivedName of the source.
operatorvarchartextderivedWho runs it.
homepagevarcharurlderivedWhere a person would look at it.
kindvarcharidderivedrouter, app, eval, measurement or catalog.
tiervarcharidderivedProvenance tier of its rows: reported, measured_by_others, measured_by_us or derived.
licensevarchartextderivedLicense of the data, as stated upstream ("none stated" is an answer).
license_urlvarcharurlderivedWhere the license is stated.
redistributionvarcharidderivedyes, attribution, unclear or no. Boards of a source marked unclear or no are not in the public board_entry; a board can override its source (board.redistribution).
citationvarchartextderivedCitation line the source asks for. Required wherever redistribution is attribution.
terms_urlvarcharurlderivedTerms of use, where they matter.
accessvarcharidderivedHow we read it: json_api, file, git, dataset, embedded_state or html.
authvarcharidderivednone, or the name of the environment variable that holds the key.
cadencevarcharidderivedHow often upstream changes: hourly, daily, weekly, monthly, irregular. We look every night.
upstream_historyvarcharidderivedWhat upstream itself keeps: none, rolling or full.
upstream_history_notevarchartextderivedThe window, where it is rolling.
notesvarchartextderivedCaveats the source module states.
boardsintegercountderivedBoards the source declares. [not null]
public_boardsintegercountderivedOf those, boards whose rows are published. [not null]
publicbooleanflagderivedTrue when at least one of its boards is published. [not null]
declaredbooleanflagderivedFalse for a source id found in the rows that no source module declares; its rows stay private. [not null]

board

Every board of every source, public or not: what it ranks, its period kind, depth, what it measures in one sentence, and the license that applies to it.

412 rows. Key: id. Also as board.csv.gz. Every source and board, public or not.

columntypeunittierdescription
idvarcharidderivedBoard id, <source_id>.<name>. [key]
source_idvarcharidderivedSource of the board. [not null; joins source.id]
titlevarchartextderivedTitle of the board.
urlvarcharurlderivedWhere a person would look at it.
entity_typevarcharidderivedWhat the board ranks.
categoryvarcharidderivedCategory or task the board is about (programming, vision), where it has one.
periodvarcharidderivedday, week, month, rolling (another trailing window), asof (a state dated by upstream) or none (no date upstream: snapshot is the only time).
metricsvarchar[]listderivedMetric names of the board; units are in board_metric.
rank_metricvarcharidderivedMetric the board's own ordering follows.
higher_is_betterbooleanflagderivedDirection of the board's metrics; a metric can override it (board_metric).
sums_to_100booleanflagderivedThe shares of one period and metric add up to 100.
depthvarchartextderivedHow much of the ranking is published: top N, or full.
measuresvarchartextderivedWhat the board measures, in one sentence.
tiervarcharidderivedProvenance tier, from the source.
licensevarchartextderivedLicense that applies: the board's own where it has one, else the source's.
license_urlvarcharurlderivedWhere that license is stated.
redistributionvarcharidderivedyes, attribution, unclear or no: the board's own where it has one, else the source's.
citationvarchartextderivedCitation line that applies.
notesvarchartextderivedCaveats the source module states.
publicbooleanflagderivedTrue when redistribution is yes or attribution: the board's rows are in the public board_entry. [not null]
declaredbooleanflagderivedFalse for a board id found in the rows that no source module declares; its rows stay private. [not null]

board_metric

One row per board and metric: the unit of board_entry.value for that metric and which direction is better.

833 rows. Key: board_id, metric. Also as board_metric.csv.gz. Every source and board, public or not.

columntypeunittierdescription
board_idvarcharidderivedBoard the metric belongs to. [key; joins board.id]
metricvarcharidderivedMetric name, as in board_entry.metric. [key]
unitvarcharidderivedUnit of the value, from the closed list of boards/contract.py. [not null]
higher_is_betterbooleanflagderivedThe metric's own direction where declared, else the board's.
descriptionvarchartextderivedWhat the metric is, where the source says.

board_coverage

One row per board, public or not: rows held, rows whose entity is mapped to a canonical id, first and last period, first and last snapshot, last collection time. Counts of private boards are metadata and are listed; their rows are not.

405 rows. Key: board_id. Also as board_coverage.csv.gz. Every source and board, public or not.

columntypeunittierdescription
board_idvarcharidderivedBoard the row counts. [key; joins board.id]
source_idvarcharidderivedSource of the board. [not null; joins source.id]
rowsbigintcountderivedRows of board_entry. [not null]
mappable_rowsbigintcountderivedRows whose entity type has an entity table (residual rows and types without a table cannot be mapped). [not null]
mapped_rowsbigintcountderivedOf those, rows with an entity_id. [not null]
entitiesbigintcountderivedDistinct entity_raw. [not null]
mapped_entitiesbigintcountderivedDistinct entity_id. [not null]
first_period_enddatedatederivedOldest period_end; null on boards with no date of their own.
last_period_enddatedatederivedNewest period_end.
first_snapshotdatedatederivedOldest snapshot kept. [not null]
last_snapshotdatedatederivedNewest day the board was collected (max of last_seen). [not null]
last_collected_attimestamptimestampderivedUTC time of that collection: the board's last sync. [not null]
publicbooleanflagderivedThe board's rows are published. [not null]

name_map

How every published name was resolved: one row per source, entity type, raw name, provider and join key, with the canonical ids boards/names.py gives it. board_entry.entity_id and provider_id are this table joined to the rows at build time. Names of private boards are in the private copy only.

26,157 rows. Key: source_id, entity_type, entity_raw, provider_raw, join_key. Also as name_map.csv.gz.

columntypeunittierdescription
source_idvarcharidderivedSource whose spelling this is. [key; joins source.id]
entity_typevarcharidderivedEntity type of the rows. [key]
entity_rawvarchartextderivedName as published. [key]
provider_rawvarchartextderivedProvider as published, on model by provider rows. [key]
join_keyvarcharjsonderivedThe ids the source publishes next to the name and boards/names.py joins on (from board_entry.extra), as a JSON object; null where the source has none. [key]
entity_idvarcharidderivedCanonical id, null when unmapped. [joins model.id]
provider_idvarcharidderivedCanonical provider id, null when unmapped or no provider. [joins provider.id]
variant_from_namevarchartextderivedVariant the mapping took off the name (free, a checkpoint date).
methodvarchartextderivedHow the entity resolved: alias, seed, crosswalk, id, name; else why not: no alias, ambiguous, id differs, no entity table (the entity type has none), proposed (the alias points at an id no entity table holds yet).
provider_methodvarchartextderivedThe same for the provider.
proposed_idvarcharidderivedThe id an alias points at when no entity table holds it yet (method = proposed).
rowsbigintcountderivedRows of board_entry with this name.

app

Canonical apps: OpenRouter's app slug where it has one, else the title as a slug. Apps are products, never people.

162 rows. Key: id. Also as app.csv.gz.

columntypeunittierdescription
idvarcharidderivedCanonical id: lowercase, stable, never reused. [key]
display_namevarchartextderivedName for people.
homepagevarcharurlderivedHome page, where known.
sourcevarcharidderivedWhere the row was seeded from.
addeddatedatederivedDay the row was added to the curated file.

app_alias

Every spelling of an app that a source uses.

172 rows. Key: file, line. Also as app_alias.csv.gz.

columntypeunittierdescription
filevarcharidderivedFile under data/curated/aliases/ the row comes from. <source_id>.csv is the source's own file; _seed_<origin>.csv is a crosswalk found in raw files. [key]
lineintegercountderivedLine of that file. [key]
source_idvarcharidderivedSource whose spelling this is. [joins source.id]
raw_namevarchartextderivedApp name as that source writes it.
entity_idvarcharidderivedCanonical id the name maps to. An id that no entity table holds yet is a proposal and is not applied to board_entry. [joins app.id]
variantvarchartextderivedVariant the name implies (free, a checkpoint date), copied to board_entry.variant where the source gave none.
methodvarcharidderivedHow the mapping was made: exact, rule, manual, crosswalk.
addeddatedatederivedDay the row was added to the curated file.
openrouter_app_idvarchartextderivedColumn of data/curated/app_alias that the schema does not declare yet; passed through as text.
openrouter_slugvarchartextderivedColumn of data/curated/app_alias that the schema does not declare yet; passed through as text.

family

Model families (meta/llama-3.1): the target of family aliases, for boards that name a family and not one model (Ollama pulls a library entry, not a checkpoint). A model row that no model alias answers may carry a family id in entity_id, marked family in name_map.method.

206 rows. Key: id. Also as family.csv.gz.

columntypeunittierdescription
idvarcharidderivedCanonical family id, lab/family. [key]
display_namevarchartextderivedName for people.
lab_idvarcharidderivedLab of the family. [joins lab.id]
sourcevarcharidderivedWhere the row was seeded from.
addeddatedatederivedDay the row was added to the curated file.

family_alias

Every spelling of a family that a source uses, from data/curated/aliases/family/.

206 rows. Key: file, line. Also as family_alias.csv.gz.

columntypeunittierdescription
filevarcharidderivedFile under data/curated/aliases/ the row comes from. <source_id>.csv is the source's own file; _seed_<origin>.csv is a crosswalk found in raw files. [key]
lineintegercountderivedLine of that file. [key]
source_idvarcharidderivedSource whose spelling this is. [joins source.id]
raw_namevarchartextderivedFamily name as that source writes it.
entity_idvarcharidderivedCanonical id the name maps to. An id that no entity table holds yet is a proposal and is not applied to board_entry. [joins family.id]
variantvarchartextderivedVariant the name implies (free, a checkpoint date), copied to board_entry.variant where the source gave none.
methodvarcharidderivedHow the mapping was made: exact, rule, manual, crosswalk.
addeddatedatederivedDay the row was added to the curated file.

gateway

Canonical gateways and routers that sit in front of providers (OpenRouter, Vercel AI Gateway, Cloudflare AI Gateway): the entities of gateway benchmark boards such as ComputeSDK.

27 rows. Key: id. Also as gateway.csv.gz.

columntypeunittierdescription
idvarcharidderivedCanonical gateway id. [key]
display_namevarchartextderivedName for people.
homepagevarcharurlderivedHome page of the gateway.
kindvarcharidderivedgateway or router, as the registry files it.
sourcevarcharidderivedWhere the row was seeded from.
addeddatedatederivedDay the row was added to the curated file.

gateway_alias

Every spelling of a gateway that a source uses, from data/curated/aliases/gateway/.

15 rows. Key: file, line. Also as gateway_alias.csv.gz.

columntypeunittierdescription
filevarcharidderivedFile under data/curated/aliases/ the row comes from. <source_id>.csv is the source's own file; _seed_<origin>.csv is a crosswalk found in raw files. [key]
lineintegercountderivedLine of that file. [key]
source_idvarcharidderivedSource whose spelling this is. [joins source.id]
raw_namevarchartextderivedGateway name as that source writes it.
entity_idvarcharidderivedCanonical id the name maps to. An id that no entity table holds yet is a proposal and is not applied to board_entry. [joins gateway.id]
variantvarchartextderivedVariant the name implies (free, a checkpoint date), copied to board_entry.variant where the source gave none.
methodvarcharidderivedHow the mapping was made: exact, rule, manual, crosswalk.
addeddatedatederivedDay the row was added to the curated file.

hardware

Canonical accelerators and systems (nvidia-h200, nvidia-b300): the entities of model by hardware boards (MLPerf) and the target of board_entry.hardware_id.

21 rows. Key: id. Also as hardware.csv.gz.

columntypeunittierdescription
idvarcharidderivedCanonical hardware id. [key]
display_namevarchartextderivedName for people.
vendorvarcharidderivedVendor slug.
kindvarcharidderivedaccelerator or system.
sourcevarcharidderivedWhere the row was seeded from.
addeddatedatederivedDay the row was added to the curated file.

hardware_alias

Every spelling of a hardware that a source uses, from data/curated/aliases/hardware/.

31 rows. Key: file, line. Also as hardware_alias.csv.gz.

columntypeunittierdescription
filevarcharidderivedFile under data/curated/aliases/ the row comes from. <source_id>.csv is the source's own file; _seed_<origin>.csv is a crosswalk found in raw files. [key]
lineintegercountderivedLine of that file. [key]
source_idvarcharidderivedSource whose spelling this is. [joins source.id]
raw_namevarchartextderivedHardware name as that source writes it.
entity_idvarcharidderivedCanonical id the name maps to. An id that no entity table holds yet is a proposal and is not applied to board_entry. [joins hardware.id]
variantvarchartextderivedVariant the name implies (free, a checkpoint date), copied to board_entry.variant where the source gave none.
methodvarcharidderivedHow the mapping was made: exact, rule, manual, crosswalk.
addeddatedatederivedDay the row was added to the curated file.

lab

Canonical labs: the part of the OpenRouter slug before the slash.

430 rows. Key: id. Also as lab.csv.gz.

columntypeunittierdescription
idvarcharidderivedCanonical id: lowercase, stable, never reused. [key]
display_namevarchartextderivedName for people.
homepagevarcharurlderivedHome page, where known.
sourcevarcharidderivedWhere the row was seeded from.
addeddatedatederivedDay the row was added to the curated file.

lab_alias

Every spelling of a lab that a source uses.

87 rows. Key: file, line. Also as lab_alias.csv.gz.

columntypeunittierdescription
filevarcharidderivedFile under data/curated/aliases/ the row comes from. <source_id>.csv is the source's own file; _seed_<origin>.csv is a crosswalk found in raw files. [key]
lineintegercountderivedLine of that file. [key]
source_idvarcharidderivedSource whose spelling this is. [joins source.id]
raw_namevarchartextderivedLab name as that source writes it.
entity_idvarcharidderivedCanonical id the name maps to. An id that no entity table holds yet is a proposal and is not applied to board_entry. [joins lab.id]
variantvarchartextderivedVariant the name implies (free, a checkpoint date), copied to board_entry.variant where the source gave none.
methodvarcharidderivedHow the mapping was made: exact, rule, manual, crosswalk.
addeddatedatederivedDay the row was added to the curated file.

model

Canonical models. id is the OpenRouter slug where one exists, else the lowercased Hugging Face repo id, else lab/name minted by hand. Thinking mode, effort, agent, free tier, quantization and checkpoint dates are never in the id.

1,896 rows. Key: id. Also as model.csv.gz.

columntypeunittierdescription
idvarcharidderivedCanonical model id. [key]
display_namevarchartextderivedName for people.
lab_idvarcharidderivedLab that built it. [joins lab.id]
open_weightsbooleanflagderivedTrue where a catalog names a Hugging Face repository; null is unknown, not false.
hf_repovarcharidderivedHugging Face repository.
familyvarcharidderivedModel family, a convenience, not a key.
releaseddatedatederivedRelease day (the day OpenRouter listed it, where that is the origin).
sourcevarcharidderivedWhere the row was seeded from.
addeddatedatederivedDay the row was added to the curated file.
modalityvarcharidderivedWhat the model takes and makes: text, image, video, audio, embedding, or a combination as the catalog states it.

model_alias

Every spelling of a model that a source uses, and the canonical model it maps to. Curated as CSV under data/curated/aliases/model/ so a change is a reviewable diff.

9,714 rows. Key: file, line. Also as model_alias.csv.gz.

columntypeunittierdescription
filevarcharidderivedFile under data/curated/aliases/ the row comes from. <source_id>.csv is the source's own file; _seed_<origin>.csv is a crosswalk found in raw files. [key]
lineintegercountderivedLine of that file. [key]
source_idvarcharidderivedSource whose spelling this is. [joins source.id]
raw_namevarchartextderivedModel name as that source writes it.
entity_idvarcharidderivedCanonical id the name maps to. An id that no entity table holds yet is a proposal and is not applied to board_entry. [joins model.id]
variantvarchartextderivedVariant the name implies (free, a checkpoint date), copied to board_entry.variant where the source gave none.
methodvarcharidderivedHow the mapping was made: exact, rule, manual, crosswalk.
addeddatedatederivedDay the row was added to the curated file.
kilo_slugvarchartextderivedColumn of data/curated/model_alias that the schema does not declare yet; passed through as text.
openrouter_idvarchartextderivedColumn of data/curated/model_alias that the schema does not declare yet; passed through as text.
aa_slugvarchartextderivedColumn of data/curated/model_alias that the schema does not declare yet; passed through as text.
openrouter_permaslugvarchartextderivedColumn of data/curated/model_alias that the schema does not declare yet; passed through as text.
da_model_idvarchartextderivedColumn of data/curated/model_alias that the schema does not declare yet; passed through as text.
vercel_idvarchartextderivedColumn of data/curated/model_alias that the schema does not declare yet; passed through as text.
originvarchartextderivedColumn of data/curated/model_alias that the schema does not declare yet; passed through as text.
aa_idvarchartextderivedColumn of data/curated/model_alias that the schema does not declare yet; passed through as text.

provider

Canonical inference providers: OpenRouter's provider slug. The provider column of the older tables uses the same ids.

188 rows. Key: id. Also as provider.csv.gz.

columntypeunittierdescription
idvarcharidderivedCanonical id: lowercase, stable, never reused. [key]
display_namevarchartextderivedName for people.
homepagevarcharurlderivedHome page, where known.
sourcevarcharidderivedWhere the row was seeded from.
addeddatedatederivedDay the row was added to the curated file.

provider_alias

Every spelling of a provider that a source uses.

700 rows. Key: file, line. Also as provider_alias.csv.gz.

columntypeunittierdescription
filevarcharidderivedFile under data/curated/aliases/ the row comes from. <source_id>.csv is the source's own file; _seed_<origin>.csv is a crosswalk found in raw files. [key]
lineintegercountderivedLine of that file. [key]
source_idvarcharidderivedSource whose spelling this is. [joins source.id]
raw_namevarchartextderivedProvider name as that source writes it.
entity_idvarcharidderivedCanonical id the name maps to. An id that no entity table holds yet is a proposal and is not applied to board_entry. [joins provider.id]
variantvarchartextderivedVariant the name implies (free, a checkpoint date), copied to board_entry.variant where the source gave none.
methodvarcharidderivedHow the mapping was made: exact, rule, manual, crosswalk.
addeddatedatederivedDay the row was added to the curated file.

provider_model_alias

A provider's native model id mapped to the canonical model: the crosswalk that resolves price and speed boards, where a row is a model on a provider.

9,264 rows. Key: file, line. Also as provider_model_alias.csv.gz.

columntypeunittierdescription
filevarcharidderivedFile under data/curated/aliases/provider_model/. [key]
lineintegercountderivedLine of that file. [key]
source_idvarcharidderivedSource that states the mapping. [joins source.id]
providervarchartextderivedProvider, as that source writes it.
provider_model_idvarchartextderivedThe provider's own id for the model.
entity_idvarcharidderivedCanonical model id. [joins model.id]
methodvarcharidderivedHow the mapping was made.
addeddatedatederivedDay the row was added to the curated file.
variantvarchartextderivedColumn of data/curated/provider_model_alias that the schema does not declare yet; passed through as text.
hf_repovarchartextderivedColumn of data/curated/provider_model_alias that the schema does not declare yet; passed through as text.
statusvarchartextderivedColumn of data/curated/provider_model_alias that the schema does not declare yet; passed through as text.

task

Canonical task categories (agent:memory_extraction, code-debugging): the entities of task boards such as OpenRouter's task classifications.

29 rows. Key: id. Also as task.csv.gz.

columntypeunittierdescription
idvarcharidderivedCanonical task id. [key]
display_namevarchartextderivedName for people.
groupvarcharidderivedTask group (agent, code, data).
sourcevarcharidderivedWhere the row was seeded from.
addeddatedatederivedDay the row was added to the curated file.