AI ROI Calculator

August 17, 2026 · View on GitHub

Built by OptimNow. Work out whether an AI project pays for itself, with live model prices, the operational costs everyone forgets, and arithmetic you can audit line by line.

Open the calculator · Methodology

Overview

Most AI business cases are built on two numbers: a token price copied off a pricing page, and a saving someone estimated in a meeting. Both are usually wrong, and neither survives contact with a CFO.

Token prices move, and the published rate is rarely what you pay once prompt caching and batch processing enter the picture. The model itself is often the smaller half of the bill: orchestration, retrieval, monitoring and guardrails run on every request and rarely appear in the estimate at all. On the value side, "productivity gains" is not a number a finance team can put in a plan.

This calculator uses a 3-layer framework that matches how production AI systems are billed, and shows its working at every step.

The three layers

Layer 1, infrastructure. Model inference, priced from the live OptimToken catalog: input and output rates plus the provider's published prompt-cache read and batch prices. Supports multi-model routing, flat per-request contracts, and retries (a retried call is charged again, so it belongs here).

Layer 2, harness. Orchestration, retrieval and vector search, tool APIs, logging, guardrails, egress and storage. Entered per 1,000 calls, which is how vendors quote them.

Layer 3, business value. 4 archetypes: cost displacement, revenue uplift, retention uplift and premium monetization. Each carries a realization rate, because a technically successful output does not automatically become money.

What you get out

ROI, net monthly benefit, payback, unit cost and break-even volume, plus a cumulative-profit curve, a tornado chart ranking which assumption moves ROI most, and unlimited saved scenarios for side-by-side comparison. Every formula is documented in METHODOLOGY.md, so stakeholders can check the arithmetic instead of trusting it.

Model prices stay current

Prices come from the OptimToken catalog, which tracks 250+ models and refreshes daily. Pick a model from the dropdown and its rates fill in, including the cache and batch prices its provider publishes.

The catalog resolves in order: a browser cache under 24h old, then the live API, then a stale cache, then a snapshot embedded in the app. The interface states which layer answered and the date the prices were published, so a figure never appears without its provenance.

Models are repriced automatically whenever the catalog resolves. Two exceptions are deliberate: a model you priced by hand keeps your negotiated rate, and a scenario loaded from storage keeps the prices it recorded, because it is the record of a decision made on a date.

Arriving from the hub with ?useCase=&volume=&model=&batch= preloads the scenario, so a comparison there continues here without retyping.

Getting started

npm install
npm run dev        # http://localhost:3000
npm run build      # production build to dist/
npm run preview    # serve the production build
npm test           # 114 tests

Requires Node.js 20+. No API keys, no backend, no account.

To refresh the embedded price snapshot used as the offline fallback:

npm run refresh:models

A weekly GitHub Action runs the same script. It refuses to write a snapshot that drops a model any preset depends on, since that would break the app at import time.

Using it

Load a preset. 11 scenarios covering support deflection, knowledge Q&A, meeting and call summaries, marketing content, coding, invoice processing, agent workflows, e-commerce recommendations, retention and premium features. Each is bound to a real model with token counts matching the pricing hub's profiles.
Define value and scope. Use case, unit, monthly volume, then a value archetype and its drivers, then a realization rate. An equation preview shows how the three rates combine.
Set the cost model. Pick a model or enter custom pricing, set token counts, add the harness costs. Advanced mode adds multi-model routing, cache hit rate, retry rate and ops overhead.
Read the results. Two accounts are in play, and they answer different questions. Break-even volume asks how many units a month cover your costs. Payback asks how long until the setup is repaid. The in-app help works both through with a worked example.
Stress-test it. Sensitivity sliders move volume, realization, cost and value without touching your inputs. The tornado chart ranks what matters.

Project structure

ai-roi-calculator/
├─ App.tsx                       # state, layout, results panel
├─ types.ts                      # UseCaseInputs, ModelParams, Scenario
├─ constants.ts                  # 11 presets, bound to catalog models
├─ components/
│  ├─ Charts.tsx                 # ROI curve, cost/value, breakdown, tornado
│  ├─ ModelPicker.tsx            # catalog dropdown + billing basis + cost fields
│  ├─ InputComponents.tsx        # money, number, percent, section header
│  ├─ HelpGuide.tsx              # in-app guide
│  ├─ ScenarioManager.tsx        # save, load, export, import
│  ├─ ScenarioComparison.tsx     # side-by-side comparison
│  └─ ErrorBoundary.tsx
├─ utils/
│  ├─ calculations.ts            # the engine: all cost, value and ROI formulas
│  ├─ modelCatalog.ts            # live prices, cache, snapshot, repricing
│  ├─ deepLink.ts                # scenario handover from the pricing hub
│  └─ format.ts                  # shared display rules
├─ scripts/refresh-model-snapshot.mjs
├─ METHODOLOGY.md                # full mathematical specification
├─ UAT_SCENARIOS.md              # 11 acceptance scenarios
└─ DEPLOYMENT.md

Tests sit beside the code they cover, as *.test.ts.

Testing

114 tests run under Vitest, covering the calculation engine, the modules around it, scenario validation, and the shared input components:

npm test

They cover all 4 value methods, cache-read and batch pricing, per-call billing, routing blends, retry attribution, break-even and payback, deep-link parsing and validation, catalog repricing, and display formatting. Edge cases include zero volume, 0% and 100% realization, negative margins and models that have left the catalog.

The component tests render NumberInput, MoneyInput, PercentInput and SectionHeader with React Testing Library in a jsdom environment, covering min/max clamping, blur normalization of empty and negative fields, and label-to-input association.

UAT_SCENARIOS.md holds 11 manual acceptance scenarios for interface work.

Tech stack


Framework	React 19 with TypeScript in strict mode
Build	Vite
Styling	Tailwind CSS v4 via PostCSS
Charts	Recharts
Icons	Lucide React
Tests	Vitest
Hosting	Vercel

All calculation runs in the browser. Scenarios are saved to localStorage and never leave your machine. The only network call fetches the public price catalog, which sends no data about your scenario.

Analytics (GA4 and Vercel Analytics) record anonymous usage events: page views, preset loads, exports, model selections. No input values, no scenario data, no personal information.

Deployment

Push to main and Vercel deploys. DEPLOYMENT.md covers configuration, static asset troubleshooting and cache clearing.

The rest of the family


OptimToken	Compare what 250+ models cost per request, with caching and batch factored in. Source of this calculator's prices.
MCP server	The same engine as an MCP tool, so an assistant can size a business case in conversation.
cloud-finops-skills	FinOps knowledge for AI agents: AWS, Azure, GCP, AI inference, SaaS.

The MCP server consumes this repository's engine verbatim through a sync script, so both answer the same question with the same number.

Contributing

Issues and pull requests are welcome. Changes to utils/calculations.ts need a matching test and a METHODOLOGY update, and they ripple into the MCP server through its engine sync.

License

MIT. See LICENSE.

Questions about your own AI cost estimate? Talk to OptimNow.