← AI Infrastructure

LLM Endpoint Drift

Independent daily record of what changed behind a provider endpoint: which alias resolved to which model, and when behaviour shifted.

POST /v1/model/drift · 1 credit

No account needed to try any endpoint on this page.
Deterministic endpoints answer from cache in single-digit milliseconds.
Every response carries a receipt: credits charged, cache status, duration.

Free try on this page → when you need a key, Proceed to pay · API access.

Try free Proceed to pay · API access Read the spec

POST/v1/model/drift

Request

{
  "view": "aliases"
}

no request sent yet

Response

Press Run to try free — no key needed.

1endpoints

—data points banked

—time series

—median response

Why this exists

Built for hostile input

A Turkish address arrives as "Cd." or "Cad." or "Caddesi", with the flat number glued to the building number. Generic parsers drop half of it silently. This one tells you what it could not place.

Every number carries its provenance

Each archived point stores the source URL and a content hash of the page it came from, and the archive is committed to a Merkle root anchored in Bitcoin. Every commitment — with its cutoff, leaf count and anchors — is published at /proof, so you can check for yourself how current the latest one is. The bulletins themselves are public and anyone can read them; what this adds is that a number you quote can be shown to be the number we held on that date, and that it has not been edited since.

You only pay for work

Repeated inputs come from cache and cost zero credits. If a response ever fails its own schema, the call is refunded automatically.

Endpointsexpand any row to see the parameters and call it live

AI Infrastructure1

/v1/model/drift LLM Endpoint Drift 1 credit

Independent daily record of what changed behind a provider endpoint: which alias resolved to which model, and when behaviour shifted.

Providers update models behind stable endpoint names. This endpoint publishes an independent measurement: a fixed probe suite is sent daily at temperature 0, three repeats per probe, and the identity a provider declares in its own response (`modelVersion`, `system_fingerprint`, region) is recorded alongside. A change is only reported as drift when the repeats agree with each other and disagree with the previous run — same-day disagreement is noise, not drift. `aliases` answers "what was actually behind `gemini-flash-latest` on that day"; a question that cannot be answered retroactively.

Parameters

view	"aliases" \| "events" \| "probes"	`aliases`: alias → the real model behind it · `events`: drift events · `probes`: probe status from the last run. default "aliases"
provider	string	Filter by provider (gemini, groq, cerebras, mistral).
severity	"high" \| "medium" \| "low" \| "info"	Only for the `events` view.
days	integer	History window in days (only `events`). default 90
limit	integer	default 100

POST/v1/model/drift

Request

{
  "view": "aliases"
}

no request sent yet

Response

Press Run to try free — no key needed.

Pricingone key covers every endpoint

Free try on this page → when you need a key, Proceed to pay · API access.

Proceed to pay · API access All plans →

Credits are consumed per call and cache hits are free. Requests without a key run on a narrow per-IP demo quota so you can evaluate before signing up.

The API key is e-mailed the moment payment clears. Upgrades keep the same key, so your integration is never broken.

Want to finish the integration before paying?

Use it from an AI assistantthe same catalogue, exposed as MCP tools

Every endpoint is also published over the Model Context Protocol, so Claude, ChatGPT or your own agent can call it directly — no glue code, no wrapper functions. The assistant sees the same input schemas and the same metering.

Currently 68 tools. New endpoints appear automatically; nothing to update on your side.

claude_desktop_config.json

{
  "mcpServers": {
    "temsor": {
      "url": "https://api.temsor.com/mcp",
      "headers": { "x-api-key": "fk_live_…" }
    }
  }
}

Quickstart

Official client: npm install temsor-api · npmjs.com/package/temsor-api

Node.js

import Client from "temsor-api";

const api = new Client({ apiKey: KEY });
const r = await api.${(() => {
  const first = apis[0];
  return first.slug.split(/[/-]/).map((p, i) => (i === 0 ? p : p[0].toUpperCase() + p.slice(1))).join('');
})()}(${esc(JSON.stringify(apis[0].examples[0]?.input ?? {}))});
console.log(r, r.$meta);

cURL

curl -s https://api.temsor.com/v1/model/drift \
  -H "content-type: application/json" \
  -H "x-api-key: fk_live_…" \
  -d '{"view":"aliases"}'

JavaScript

const res = await fetch("https://api.temsor.com/v1/model/drift", {
  method: "POST",
  headers: { "content-type": "application/json", "x-api-key": KEY },
  body: JSON.stringify({"view":"aliases"})
});
const { data } = await res.json();

Python

import httpx
r = httpx.post("https://api.temsor.com/v1/model/drift",
    headers={"x-api-key": KEY},
    json={"view":"aliases"})
data = r.json()["data"]