TokenGauge
FREE WORKFLOW COST LEDGER

See which workflow moved the AI bill.

Attribute aggregate token usage to projects and features, apply the exact provider rate card, and compare cost per accepted answer. Import or export a transparent canonical CSV; everything runs locally in your browser.

Build the ledgerReconcile one provider bill

Rate snapshot verified 2026-08-16 · no login · no usage upload

PROJECT / WORKFLOW ATTRIBUTION

Replace one total
with accountable rows.

Provider reports group usage differently and do not know your application's feature names. Add project and workflow labels when you log requests, aggregate the token fields, then map each row to the exact rate band here.

Replace the sample rows with aggregate workflow totals.

Use one row per project, workflow, and exact price band. The values never leave this page.

Canonical CSV schema and provider-report limits

project,workflow,rate_card_id,input_tokens,cached_input_tokens,output_tokens,attempts,accepted_answers

OpenAI organization usage can group by project, user, API key, model, batch, and service tier. Anthropic can group by workspace, API key, model, service tier, and context window. Neither report supplies your application's workflow label, so add that label in your own request log before aggregating. TokenGauge does not pretend the provider exports are interchangeable.

WORKFLOW 01
WORKFLOW 02
WORKFLOW COST LEDGER

$8.32 modeled token spend

Support replies is the largest row at 78.4% of modeled spend.

Accepted-answer rate88.0%1,320 accepted / 1,500 attempts
Cost / accepted answer$0.0063Across every entered workflow
Non-accepted attempt estimate$1.09Uniform-attempt approximation
BY WORKFLOW
Support repliesExample SaaS · OpenAI GPT-5.6 Terra
$6.52$0.0077 / accepted
Document extractionExample SaaS · Anthropic Claude Haiku 4.5
$1.80$0.0038 / accepted
BY PROJECT
Example SaaS100.0% of modeled spend
$8.32

These are rate-card estimates, not an invoice. Map each aggregate to the price band that covered its individual requests. Tools, cache writes/storage, media, regional or priority uplifts, taxes, credits, and provider rounding remain outside rows that do not explicitly represent them.

NEXT STEPTest the largest controllable workflow, not the smallest sticker price.

Reconcile the provider bill first, then use a bounded method and the same acceptance rule on both variants.

NEED THE ROWS GENERATED AUTOMATICALLY?

Fixed £75 attribution setup for one workflow.

I can instrument one authorized Node.js/TypeScript or Python codebase, add project/workflow and retry-aware usage fields, export this ledger schema, and verify it with focused tests. Read the implementation model first →

See the exact scope
WHY THE SCHEMA IS EXPLICIT

Portable does not mean provider-blind.

01

OpenAI grouping

The organization Usage API can group completions by project, user, API key, model, batch, and service tier. Its Costs endpoint is the financial reconciliation source.

Official Usage API ↗
02

Anthropic grouping

The Messages Usage Report separates uncached input, cache creation, cache reads, and output, with workspace, API-key, model, tier, and context dimensions.

Official usage report ↗
03

Application attribution

OpenTelemetry defines provider, model, workflow, input, output, cache-read, and reasoning attributes, but cost and retry metrics remain an active standardization problem.

Official GenAI attributes ↗