llm-usage

Auto-published, anonymized, aggregate LLM usage statistics across providers — a public dataset that powers a personal portfolio page.

Today this covers OpenAI and Anthropic. The schema is provider-agnostic and extends cleanly to additional providers (Google, Mistral, xAI, local models, etc.) as they are added — see Extensibility.

What this is

This repo is a small, machine-readable view of how much LLM usage flows through my tools and projects, rolled up to safe aggregates. It is regenerated automatically and committed here so a static portfolio page can fetch it without a backend.

What this is NOT

There is no PII, no prompts, no completions, no account identifiers, no API keys, and no per-request content in this repo. Every file contains only counts, sums, and averages computed over a time window. Nothing here can be traced back to an individual conversation, user, or credential.

File layout

The dataset is published as a set of pre-aggregated JSON files, one per rolling window:

File	Window
usage.json	Default view (365 days)
usage-7.json	Last 7 days
usage-30.json	Last 30 days
usage-90.json	Last 90 days
usage-365.json	Last 365 days (canonical full)

usage.json is an alias of the default (365-day) window so consumers can fetch a single stable file without picking a period.

Fetch URL

The canonical, full-history file:

https://raw.githubusercontent.com/aneym/llm-usage/main/usage-365.json

Swap the filename for usage-7.json, usage-30.json, or usage-90.json to fetch a narrower window, or usage.json for the default view.

Update cadence

The files are refreshed roughly every ~15 minutes by an automated publisher that pulls each provider's usage/billing API, anonymizes and aggregates the results, and commits the regenerated JSON. generatedAt inside each file records the exact moment a snapshot was produced.

JSON schema (top level)

Each usage-*.json file is a single object with these keys:

Key	Type	Description
period	string	The window this file covers (e.g. "365", "30").
generatedAt	string	ISO-8601 timestamp of when this snapshot was produced.
source	string	Identifier of the publisher/pipeline that generated the file.
totals	object	Aggregate cost, token, request, latency, and success-rate figures for the window.
daily	object[]	Per-day aggregates: date, cost, tokens, requests, and the top model that day.
byModel	object[]	Per-model breakdown: requests, token splits, and cost, with provider attribution.
byProvider	object	Per-provider rollups (openai, anthropic, …) of requests, cost, and tokens.
trends	object[]	Compact time series for charts: { t, cost, tokens, requests } points.

Full field-by-field definitions live in SCHEMA.md.

Extensibility

byProvider and byModel are open maps/lists, not fixed shapes. Adding a provider means the publisher emits a new key under byProvider and new entries in byModel (each tagged with its provider) — no schema change required for consumers. Charts and totals are computed across whatever providers are present, so the dataset generalizes to any number of LLM vendors.