Sunk Cost Free

What it does

Sunk Cost is a free, single-page web calculator that answers a specific question: if you buy hardware to run large language models locally, how long until that purchase pays for itself compared to just paying for a cloud API? It was built after its creator kept hearing "just buy a Mac and run models locally, it pays for itself" and wanted actual numbers.

Key features

You enter a machine (memory, power draw, memory bandwidth), a model, and your expected daily usage (tokens per day, context window), and the calculator estimates local inference throughput, electricity cost, and the break-even point against equivalent API pricing. It also includes a model capability leaderboard that ranks local models against frontier models like Claude and GPT, a hardware pricing database, and buying-guide style comparisons. In practice the tool tends to surprise people — commenters on its Hacker News launch reported break-even periods of decades for some local setups, though they noted privacy, data sovereignty, fine-tuning ability and avoiding vendor lock-in can justify the cost regardless of payback time.

Pricing

Free, with no accounts, subscriptions or paid tiers. The site states it stores no cookies and no IP address — only the request's country — once you stop typing your inputs.

Works with

Sunk Cost is a standalone browser calculator, not a coding agent, and it doesn't load Agent Skills, MCP servers or an AGENTS.md file. It's meant as a planning tool you use before deciding whether to buy hardware for running local models with tools like Claude Code, Cursor, Ollama or LM Studio, not as an integration into any of them.

How to install

Nothing to install — open sunkcost.ai in a browser and fill in the calculator fields.

Who should use it

Developers and teams weighing whether to buy a workstation or Mac to run coding models locally instead of paying per-token for a hosted API, and anyone who wants to sanity-check the popular "local models pay for themselves" claim against real hardware and electricity numbers.