ChangeCost.dev

Public alpha · Unreleased research preview

Build with AI.
Budget with
evidence.

FinOps for AI-assisted software changes.

Explore an early local CLI for cost evidence, offline estimates and reconciliation. The source is public so you can inspect the work, try the documented experiments and follow what comes next.

Explore the alpha on GitHub Explore the approach

An unreleased source preview. Broader provider imports and forecast calibration remain work in progress.

Change overviewIllustrative workflow
UNIT OF WORK / SOFTWARE CHANGE

Add a new capability

Implementation → refinement → verification

Estimated costUnpricedComparable history needed
Provider chargesUnknownBilling evidence needed
Attempt coverageALL WORK COUNTS
  1. Initial implementationThe first approachIncluded in scope
  2. Retry & refinementThe work between versionsIncluded in scope
  3. Abandoned approachSpend that still mattersIncluded in scope
Evidence incompleteUnknown ≠ zero

A clear view of the evidence. Room for what’s unknown.

Built around clear principlesLocal-first by design Explainable estimates Public alpha source

One cost question. Every scale.

Your next change deserves clarity.

From your own side project to an organisation’s AI development practice.

Individual developers

Understand the work behind your usage. Bring more evidence to your next AI coding decision.

Share your workflow

Startups

Make AI spend part of the build decision. Explore what retries and changing approaches mean for your runway.

Discuss your use case

Small & medium businesses

Connect development choices with cost questions. Give engineering and business teams a shared vocabulary.

Bring your cost questions

Enterprise teams

Explore traceable evidence and explicit cost assumptions for your AI development practice.

Discuss evidence needs

We welcome these perspectives. The alpha is local and experimental; hosted team management and enterprise controls are outside its scope.

01 / The cost question

A token count is only
part of the story.

A change can span multiple sessions, revisions and discarded approaches. Understanding its economics means accounting for the work it took to get there.

ChangeCost puts the software change at the centre: what you expected to spend, what the evidence shows, and what remains unknown.

02 / Capabilities & direction

A feedback loop for smarter spending.

Partial alpha implementation
  1. 01

    Quote

    Experiment with offline estimates from complete synthetic history.

  2. 02

    Bind

    Explicitly connect runs and evidence to a change in the local CLI.

  3. 03

    Meter

    Inspect selected local sources. Coverage and accounting authority remain limited.

  4. 04

    Settle

    Reconcile normalized evidence offline. Keep charges, estimates and unknowns distinct.

  5. 05

    Calibrate

    Planned: evaluate predictions against real outcomes before claiming reliability.

The public source includes explicit records, normalized reconciliation and synthetic-history quotes. This feedback loop is not a claim of complete provider coverage or validated real-world predictions. Read the current scope and limits →

03 / Token economics

Token economics.
Built into the approach.

Consumption, notional value and money paid answer different questions. The planned approach keeps them separate.

USAGE

What was consumed?

Input, output and cache token categories where evidenced, plus provider credits with their own units. A usage counter alone does not establish a cash charge.

ESTIMATED COST

What might it cost?

An estimate with its assumptions, billing basis and uncertainty. It must remain identifiable as an estimate.

BILLED CHARGES

What was charged?

Provider-reported money, where available. Subscription allowances and missing billing evidence need explicit treatment.

04 / Evidence before confidence

A useful estimate
knows its limits.

The pilot will evaluate estimates against completed work, using only evidence available when each quote was made.

  • Show the comparable history behind a quote.
  • Include retries and abandoned attempts.
  • Return “unpriced” when evidence is insufficient.
  • Keep missing amounts distinct from observed zero.
Read the methodology

05 / Where we are

Public alpha.
Reliability still to validate.

The public source includes a local CLI, offline quotes and reconciliation, and narrowly scoped evidence readers. Broader provider imports and forecast calibration remain work in progress. The dated project status links the source and explains the limits.

View project status

Build the evidence with us

Build better cost decisions.
Start with your next change.

Building solo or shaping an organisation’s AI practice? Help us understand the cost decisions you need to make.

Discuss the pilot

Please don’t send prompts, credentials or private usage exports.