One day of OpenRouter traffic. Each model's tokens are split by the apps that call it and by the providers that served it. OpenRouter is the only router that publishes the provider layer, so only its branch reaches the provider layer.
OpenRouter is the only router that publishes token counts. Vercel AI Gateway publishes each model's share and no total, so its branch is an estimate: an assumed total, from Vercel's own public statements, times the published shares. It is hatched and marked est., and the card behind that mark gives the sources. TrustedRouter publishes its own measurements of providers and no usage, so it is not drawn.
Each model page lists its top 5 apps. That mix is scaled to the model's daily tokens. Everything else is "Other and unnamed apps".
The per-model provider charts cover about a quarter of the providers-page total, so read shares, not totals. OpenRouter excludes private and zero-retention traffic at the source.
Click a model to open its provider board. Click an app, a provider or a legend entry to isolate its flow. Esc clears.
OpenRouter's top 20 apps for the week. An app names itself with a request header. Traffic without one is unnamed, and only OpenRouter publishes apps at all.
Tokens of the top 20 apps divided by OpenRouter's tokens for the same week.
OpenRouter's own labels. An app with two labels is split evenly between them.
Kilo Code and OpenCode publish model rankings from their own users. Those are whole-app figures, whichever router or key the traffic used, so they overlap with OpenRouter's count of the same apps and are never added to it.
The ChatGPT and Claude.ai task mixes are different users under different taxonomies. They sit in a fold and are not comparable with the app categories.
Two coding agents publish their own usage. Tokens by model lab, each bar its own 100%.
OpenRouter publishes token counts for everything. Vercel AI Gateway publishes daily shares by model and lab, and two top-10 rankings of apps and providers. TrustedRouter publishes its own measurements of providers. Every other router we know of is in the registry below, listed and not collected. Nothing is summed across routers.
Both routers publish a listed input price for each model on each provider. Where the same model is served by the same provider on both, the two quotes sit side by side. OpenRouter's is the provider's plain endpoint where it has one; a flex or batch tier is a different product and is used only if nothing else exists. These are quotes, not bills.
OpenRouter and Vercel follow the window control. Vercel's figure for a window is the mean of its daily shares, a model missing from a day's list counting as zero that day, because Vercel publishes no totals to weight the days by.
One bar per router: its tokens by the lab that built the model, same labs and colours in every bar, and the same in the Apps section. TrustedRouter publishes no usage, so it has no bar.
OpenRouter per model through its Hugging Face id. Vercel by lab.
Tokens by model lab, one bar per router. Each bar is that router's own 100%.
OpenRouter's top 20 models of the week. Shares on Vercel and OpenCode are matched by model name. [low] means the model is below that router's published top list.
Listed input price per million tokens, lowest to highest across the model's endpoints on OpenRouter.
Click a row. Effective price is what OpenRouter reports customers paid after cache reads and discounts over 30 days. Cache counts are asserted by providers. The Compare tab sets them against our own probe.
In the price and speed chart, the line joins providers that no other provider beats on both.
Quoted against effective input price. Below the dashed line, caching brought the bill under the quote.
Effective input price against speed. The line joins providers nobody beats on both.
OpenRouter's providers page, rolling day, all models. By dollars = tokens times the blended effective price of that model at that provider, an estimate.
Quote against builder: the median, across the top 20 models, of the provider's lowest listed input price against the builder's own endpoint. Bill against quote: token-weighted effective price over listed price, 30 days.
From OpenRouter, over the provider's endpoints on the top models: token-weighted cache hit over 30 days, and the median tokens per second, first-token time and uptime of the latest day.
First-token time and availability from TrustedRouter's own monitor regions, all models pooled, with its sample count. Not by OpenRouter, not by the providers, not by us.
Daily share by provider, top eight and everyone else.
For one model, does a dearer provider answer faster? One dot per model: the rank correlation between its providers' price and their time to first token.
No router or provider publishes which chip served a request. This section shows what does exist and marks the rest [x].
Declared per endpoint on OpenRouter, for the top 20 models. "Not declared" is the provider leaving the field empty.
Three providers say on their own sites that they serve on chips they design. That is their statement, linked, and not verified by anyone here.
The chip of each attested machine, from the network's public provider list.
OpenRouter tokens served by providers that run TEEs: Phala, Venice, Chutes, NEAR AI. An upper bound for Venice, which also runs plain endpoints. A lower bound on private inference overall, because OpenRouter excludes private traffic.
Verified attested: TrustedRouter verifies the provider's attestation, or the registry fully verifies an endpoint under the provider's own name. Claims TEE: neither does. Phala and Venice are claimed today, so every "all attested" number is an upper bound.
The awesome-private-inference registry checks each provider's endpoints daily: can the endpoint prove the request ran in a TEE, layer by layer.
"Verified confidential" is TrustedRouter's own verification. Its privacy tiers classify routes, one model on one provider.
Chutes meters dollars per chute. Darkbloom's tokens per day come from the team's own posts. Neither is checked by anyone else.
Share of OpenRouter tokens served by attested providers.
What a TEE costs: each TEE endpoint's listed price, as a multiple of the same model's other providers. Left of 1× is cheaper.
Median multiple across TEE endpoints, one point per collection day.
Every provider of every panel model gets the same prompt of about 5,500 tokens twice, three seconds apart, pinned to that provider with fallbacks off. Three sweeps a day at varying hours. The first line is unique per pair, so the first request cannot hit any cache.
For both requests: the cached tokens the provider reported, the bill, and the time to the first token.
Recognized = cached tokens reported on the second request ÷ its prompt tokens. Second-request cost = its bill ÷ the first request's bill. A cell pools every pair so far and shows a provider's lowest variant of that model.
The same pair, same bytes, is also sent through TrustedRouter and Vercel AI Gateway once a day, pinned to a provider the same way, on the panel models they also serve. Each router's results are kept apart and never pooled with OpenRouter's. A provider that recognizes our repeat through one router and not through another is being reached differently, not necessarily behaving differently.
A miss is not proof of anything: the repeat can land on another replica, or the cache may be written late. A ring marks a provider that reported cached tokens on the first request, on bytes it had never seen; that is unexplained until it repeats.
We send each provider the same prompt twice, a few seconds apart, and show how much of the second one it reported as already cached.
The quoted input price of every endpoint of every panel model, every five minutes, from OpenRouter's public model pages. No key, no spend.
The model builder's own endpoint where it has one (Z.ai for GLM, Moonshot for Kimi), otherwise the median quote. 1× means "priced like the builder".
Each line is one endpoint's quoted input price as a percentage of where we first saw it. Every change the poll has seen is a step.
A cut of 1% or more fires a cache pair at the provider that cut and at the model builder in the same moment, to see whether the cache survives the repricing.
Each model's providers, placed by how their quoted price compares with the model builder's. Colour is how much of a repeated prompt they recognized.
The first turn of the coding session with no provider named, a fresh session each time, so the router chooses. Also variants sorted by price and by throughput, and five-turn sessions with and without a session_id (the router's sticky-routing key).
Which provider served each request, its price at that moment, the cached tokens and the bill.
r is the exponent in pick probability ∝ price−r, fitted by maximum likelihood with a 95% profile-likelihood interval. OpenRouter documents 2. "Turns moved" counts consecutive turns of one conversation served by different providers.
Tasks with an answer a script can check: an invoice total (one number), a lookup in a ledger, a 300-word release note with required facts and a required last sentence, and the five turns of the coding session.
No judge model. Exact number, required strings, word count, valid JSON with the right keys. Temperature 0, the same bytes for every provider.
Every provider claims to serve the same model. If one fails checks the others pass, it is serving something else: a heavier quantization, a truncated context, a broken template.
OpenRouter's cache-hit rate for that provider and model over the last 30 days: the share of all customers' prompt tokens the provider reported as cached. It depends on what customers send, not only on the provider.
Our probe: the share of an identical prompt, sent again three seconds later, that the provider reported as cached. All pairs so far, pooled. Where a provider runs several variants of a model, the lowest.
High reported and low measured: the provider's cache did not recognize an immediate repeat from us, although it reports plenty of hits overall. Low reported and high measured: its customers rarely repeat context. Neither is proof of anything at this sample size.
For each provider, the mean of its reported cache-hit rates and the mean of our measured recognition, over the models where both exist. Unweighted: each model counts once.
One line per provider from what it reports (hollow) to what we measured (filled). A long line to the left is a provider that reports much more than we saw. Providers are ordered by that gap.
A provider with one model in both sources is one observation. The table gives the number of models and of pairs behind every row.
Every table the build writes, with its rows, bytes on disk and columns, and every board of every source with its rows. Tables are counted whole; a board's bytes are its share of its source's files by rows.
First and last period are the earliest and latest day a row is about. Last sync is the newest collection. Tables with no period column are undated and rebuilt whole every build.
A source that states no license, or forbids redistribution, is collected on the same schedule but not served. Its rows are counted here and marked private.
One row per market table and per leaderboard source. The bar runs from its first period to its last; the tick is its last sync. Click a row to open it below.
How far back each source and table goes, oldest first.
One row per table: the market tables (what routers, providers and our probe publish, one Parquet file each) and the leaderboard tables (board_entry, its dimensions, entity and alias tables).
Rows, size and column count are read from the files. First and last period come from the table's own date column. Tier is the provenance of its rows: reported, measured by others, measured by us, derived.
Click a table for every column with its type, unit, tier, key and description, and a DuckDB query that reads it over HTTP.
One row per board of every source, public or not. Rows are the board's rows in board_entry: one per period, entity and metric, for each value it has had.
How the board is dated: by day, week or month, a rolling window, as of a day upstream, or none, where the snapshot is the only time.
Click a board for its metrics with units and direction, where it lives, the license and citation that apply, and how many of its rows map to a canonical id.
One row per source: who runs it, what kind of site it is, the provenance tier of its rows, its license and whether its rows may be redistributed.
Boards, rows, first and last period and last sync are summed and spanned over the source's boards, public and private.
Click a source for its homepage, terms, the citation line it asks for, how we read it and every board of it.
Reading api/v2/feed.json…
Daily searches of arXiv abstracts, arXiv links in the posts below, and a one-time literature search per topic that followed references and citations two hops out from known papers.
A small model reads each title and abstract and scores how much it bears on this market. Papers a topic search marked as essential reading are kept and marked core.
The one-line summary is the model's. The abstract is the authors'.
Reading api/v2/feed.json…
X searches naming routers and gateways together with price, fee, credit, cache and billing terms, run by month and by week from 2023.
A small model scores each post from 0 to 1 for the information it carries: numbers, measurements, outages, launches and policy changes score high; praise, questions, promotion and crypto score low. Posts at 0.7 or above are shown.
The summary is ours, written by the model, never the post's words. Likes, reposts and replies are as of collection. Open a post to read it on X; the text is not copied here.
Reading api/v2/feed.json…
Reading api/v2/feed.json…
Reading api/v2/feed.json…