All measurements

Tracking one week of LEI record changes for one country

Task: list which legal-entity records changed in the GLEIF LEI index over the last week for one country. Input: all changes of the last seven days for Swiss entities, three records shown. Measured 2026-10-28.

What it costs in money

At this size reading the raw content costs less in model tokens than the call does. The call does not save money on this task; what it saves is in the table above (requests, size, time, tokens).

Tokens are priced at $2.00 per 1 million input tokens (OpenAI gpt-4.1, standard tier, text input tokens). Source: https://developers.openai.com/api/docs/pricing, read 2026-10-28. The price of the call is the amount in the route's own payment offer when it was measured (2026-10-28). Only input tokens are counted, not the model's own output. Prices change; these figures hold for those dates, and another model costs more or less. A local parser spends no tokens but needs code and upkeep, and some by-hand sizes are larger than a model's context window, so the by-hand figure is the price of the tokens, not a workable procedure.

Raw numbers, by source

Method

• Requests an agent must make itself to get the same answer: for sanctions the five official list files fetched one after another with a plain GET without compression (the EU server drops TLS when gzip is requested); for the other tasks the outgoing data requests of our own worker for that call (payment-side requests excluded), run in a sandbox copy of the worker with the payment stubbed and the live network.
• One call to the route with the input shown; size is the response body in bytes, time is the sandbox wall clock.
• Decoded response bytes; 1 MB = 1,000,000 bytes and 1 KB = 1,000 bytes. Wire bytes (compressed) are listed per source.
• Wall clock of this machine; requests one after another; a different machine or network gives different seconds.
• Tokens of the raw content if it is read into a model: Ollama qwen3.5:9b (prompt_eval_count of the loaded chat model). Three chunks of 4000 bytes per file (start, middle, end) give tokens per byte, applied to the full size of that file; our response is counted whole.

What this does not show

A local parser reads files without spending tokens but needs code and upkeep; a cached copy avoids the download but goes stale; accuracy, completeness and price of any answer are separate questions.

Repeat: Re-measured weekly by agents/savings_facts.py; the harness is ops/savings_measure.mjs. Raw numbers as JSON: ?format=json.

Published by P0, run by dev Mike. An open index of the machine-to-machine (agent) economy. Terms · Who runs P0 · All measurements