Data API
The model fingerprint data, as a feed

YFarmX sends AI models on OpenRouter the same 50 test strings and counts the tokens each one reports back. Two models that count the same on all 50 strings share a tokenizer, which places them in one family: models built on one vocabulary. Often that is one lab's line, and some labs borrow another's: Microsoft's Phi-4 reads with OpenAI's. With other evidence it has named the maker behind two anonymous models, and both makers confirmed it. The compare tool on this site shows two models at a time. This feed hands you the whole table as JSON, with the date every figure was read.
As of 17 October 2026 the table holds 353 measured models from 59 labs, giving 72 distinct fingerprints in 59 tokenizer groups. The methodology explains how a reading becomes a verdict.
What the feed carries
| Endpoint | What comes back | Size today | Tier |
|---|---|---|---|
/api/models/catalogue | Every measured model: the provider's declared specs and pricing, the measured token counts and fingerprint, our identity verdict with its reasoning, the nearest neighbours and, where a model has been measured twice, whether its fingerprint held between the two dates. | 353 measured models, 68 of them measured twice | Developer |
/api/models/history | Every change to a model's identity or declared fields between catalogue snapshots, and every fingerprint that moved on re-measurement: what it said before, what it says now, and the date the change was seen on. | 481 events across 5 snapshots | Pro |
/api/models/clusters | Which models share a fingerprint, as clusters and the groups that fold template-boundary shifts, so you can see which names on the market share a tokenizer family. | 72 clusters, 59 groups | Pro |
/api/models/estimate | What a workload costs on every priced model: input tokens from rates fitted to real bills for each model's tokenizer family (on 18 held-out documents billed on 12 models, the input count landed 5.8% from the billed count at the median), output at list price for the answer length you set, and thinking as the think bench's median and 90th percentile with the number of runs behind them. Every line comes with its detail: the rate for each kind of text, the listing's prices and conditions, the model's limits and its bench rows. The cost estimator page shows the same table on screen for one workload at a time. | 308 priced models, 8 workload profiles | Pro |
Every figure carries its date
Each record says when its numbers were read, so a reading from last month is never mistaken for one from this morning. Five dates do the work:
| Field | Meaning | Today |
|---|---|---|
measured_at | The day a model was sent the test strings and its token counts were read. On every measured record. | 21 August 2026 to 17 October 2026 |
catalogue_date | The day the frozen catalogue snapshot behind the dataset was read from the provider. | 5 October 2026 |
declared_read | On each record, the day its declared fields (pricing, context window, tokenizer) were read: the snapshot date, or for a model listed since, the day its own listing was read. | 21 August 2026 to 17 October 2026 |
dataset_updated | The day the file itself was last rebuilt. | 17 October 2026 |
snapshot_dates | The catalogue snapshots the history compares. A change is dated to the snapshot it first appeared in. | 20 August, 21 August, 3 September, 16 September, 5 October 2026 |
Two tiers
| Developer | Pro | |
|---|---|---|
| Catalogue | Included | Included |
| History | Pro only | Included |
| Clusters | Pro only | Included |
| Cost estimate | Pro only | Included |
| Price | £29 a month | £199 a month |
Consumers normally have 14 days after entering the contract to cancel. Checkout asks for immediate access and acknowledges the cancellation rule for digital content. Your statutory rights for an ongoing service or faulty supply still apply. Read the cancellation information.
Subscribe to Developer, £29 a monthSubscribe to Pro, £199 a month
Checkout is by card through Stripe. Your key appears on screen as soon as Stripe confirms the payment, usually within seconds, and it works for as long as the subscription is paid; cancel and it stops at the end of the period. A key covers use inside your own organisation and in products that credit YFarmX. The data API terms cover billing, cancelling, refunds and what the licence allows.
Already a subscriber? Your key page opens your key again in the browser you bought it in, and Stripe's billing portal changes your card, switches plan, downloads invoices or cancels.
Pay per call in USDC
An AI agent can buy a single call with no account and no key. Ask any endpoint without a key and it answers 402 Payment Required with the price; an x402 client pays in USDC on Base and gets the data in the same request. The USDC goes from the buyer's wallet straight to ours.
| Endpoint | Price per call |
|---|---|
/api/models/catalogue | 1.00 USDC, about US$1 |
/api/models/history | 2.00 USDC, about US$2 |
/api/models/clusters | 1.00 USDC, about US$1 |
/api/models/estimate | 2.00 USDC, about US$2 |
Both x402 versions are accepted: v2 clients read the PAYMENT-REQUIREDheader and pay with PAYMENT-SIGNATURE, v1 clients read the 402 body and pay with X-PAYMENT. The receipt comes back in PAYMENT-RESPONSE(v1: X-PAYMENT-RESPONSE) with the Base transaction hash. USDC is a dollar stablecoin, so 1.00 USDC is about US$1.
If an answer is lost on the way, send the same request with the same payment header again within ten minutes. While its confirmed receipt is available, the answer is sent again at no extra charge. Automatic redelivery may become unavailable sooner. If settlement is uncertain, the reply is 503 settlement_unconfirmed: retry with the same header and follow the response's instructions. A payment header buys one request. If payment settled and our service failed to deliver, email [email protected] within 30 days with the Base transaction hash and endpoint. We will send the data, or return the USDC if we cannot. Keep the original payment header; sign a new payment only for a new purchase.
How to call it
Send the key as a bearer token. Every endpoint answers GET only.
curl -H "Authorization: Bearer YOUR_KEY" https://yfarmx.com/api/models/catalogueThe same key in your agent
The YFarmX MCP server carries four tools that read this feed: yfarmx_models_catalogue (Developer), yfarmx_models_history,yfarmx_models_clusters and yfarmx_models_estimate (Pro). Add the server with your key as the connection's bearer header and the agent can pull the table, the history and the clusters itself. Without a key the tools answer with where to get one, and the free tools on the same server keep working.
claude mcp add --transport http yfarmx https://yfarmx.com/api/mcp --header "Authorization: Bearer YOUR_KEY"catalogue takes ?lab=Google;history takes ?since=2026-09-01 and ?model=google/gemini-3.8-flash;estimate takes ?profile=code-assistant&requests=100000&effort=high, or ?chars=5400&shares=prose:0.7,code:0.3&task=code for your own mix, and &model=openai/gpt-5.5 for one model's line with the detail behind it. The full catalogue answers with an ETag, and a request carrying it back in If-None-Match gets a 304 when nothing has changed.
| Status | Body | Meaning |
|---|---|---|
| 200 | the data | Key accepted, or payment settled. |
| 304 | empty | Your copy is current. |
| 402 | the price | No key sent: pay for the call with x402, or send a key. |
| 401 | key_unknown | The key is misspelt or has been revoked. |
| 403 | tier_required | The endpoint is Pro and the key is Developer. |
| 403 | subscription_inactive | The subscription behind the key is no longer paid up: it ended, or a renewal payment failed. |
Pay by invoice instead
For an invoice, a custom plan or more than one key, tell us here.
What stays free
The compare tool, every model's record page, the MCP server and the news API stay free and open to anyone. Quoting a figure from a record page, with a link back to it, is welcome under our reuse terms. The feed exists for people who need the whole table in one fetch, kept current, with the dates attached.