UK-based independent AI efficiency projectEvidence before claims · 2026

Token Optimisation

AI efficiency · measured
Discuss a diagnostic

Subscription vs API: Compare Cost per Verified Job, Not a Token Price Guess

Subscription and API cost paths meet at matched window and accepted-job count; missing denominator keeps cost per job not calculated.

Written by

Token Optimisation

in

,

Classification: METHOD and CALCULATED scenarios only. This article does not declare that a subscription or API is cheaper for a particular workload.

Cost per Verified JobDENOMINATOR INCOMPLETE · COST UNKNOWN. MATCH PERIOD → CAPTURE USAGE → COUNT ACCEPTED → CALCULATE OR UNKNOWN. £89 is user-recorded; billing window/account-wide usage are unknown; the complete account-period job denominator remains UNKNOWN.DENOMINATOR INCOMPLETE · COST UNKNOWN1MATCH PERIOD2CAPTURE USAGE3COUNT ACCEPTED4CALCULATE OR UNKNOWN£89 is user-recorded; billing window/account-wide usage are unknown; one local result is not a full denominator.
DENOMINATOR INCOMPLETE · COST UNKNOWN. £89 is user-recorded; billing window/account-wide usage are unknown; one local result is not a full denominator.

View the full infographic → · View this article

Subscription prices are not per-million-token prices. To compare a flat monthly plan with API billing, use a common time window and denominator: accepted jobs, useful machine hours or another outcome that matters. If usage or acceptance is missing, the result is not calculated.

What is known in the current Pro 100 experiment

The recorded charge is £89 on 2 October 2026, supplied by the account owner and not invoice-verified. The billing window is unverified. A scoped local Bridge snapshot had zero local usage rows and zero verified production outputs at capture, but it does not cover other clients. The benchmark remains IN PROGRESS; those zeros are scoped observations, not an account-wide total.

Use comparable calculations

For a subscription, cost per verified job is the charge for the comparison window divided by accepted jobs in that same window. If the accepted-job count is unavailable, show “not calculated.” For an API scenario, add uncached input, cached-input and output charges using the selected model’s official price table and its currency/date. Separate retries, tools, cache writes, storage fees, taxes and discounts when relevant.

Then compare the cost against the same task class and acceptance standard. Different windows, model quality, provider allowances or human recovery work can make a simple division misleading.

Inputs that must be visible

  • Plan or model identity, price source, source date and billing period.
  • Input, cached-input and output counts as defined by that provider.
  • Requests, retries and excluded fee categories.
  • Verified-job count and the acceptance test.
  • Useful machine time, waits and human minutes if time efficiency is discussed.

The Context Cost Calculator supports official API rates and subscription measured-usage mode. Its default Pro 100 fields remain blank where measurements are not available. Do not infer subscription token rates or a plan verdict from the £89 charge alone.

More posts