A new web tool called Sunk Cost attempts to answer a question every developer running local language models eventually asks: how many API calls does it take before the GPU you already bought starts saving you money? The calculator at sunkcost.ai takes your hardware specs, your usage patterns, and current API pricing, then estimates the break-even point.
How the Calculator Works
The core comparison is straightforward. You provide your local machine's performance characteristics (or let the site estimate them from memory bandwidth and per-token read costs), your typical API usage volume and model choice, and the cost of your electricity. The tool then estimates how many tokens you need to process locally before the cumulative API savings exceed the amortized cost of your hardware.
Where direct measurement is not available, the site estimates local inference speed using a formula: memory bandwidth divided by bytes read per token. That approximation is clearly labeled whenever it is used rather than presented as a measured value. API speed affects the time comparison between local and cloud inference but does not change the cost calculation itself.
Privacy by Design
The tool is notably minimal in what it collects. Your machine speed and electricity bill data are sent only after you stop typing, with no cookies and no IP address recorded. The only geographic data captured is the country the request originated from, used to estimate regional electricity costs. Browsers configured to block tracking send nothing at all.
That privacy posture is worth noting in a category where calculator tools often quietly accumulate user data for analytics or advertising purposes. Sunk Cost treats the input as ephemeral calculation data rather than user profiling material.
What You Can Change
The assumptions behind the calculation are adjustable. Users can modify their API provider, model selection, usage volume, hardware cost, and electricity rate to see how the break-even point shifts under different conditions. That flexibility matters because the answer varies dramatically depending on whether you are comparing against a cheap model like GPT-4o-mini at fractions of a cent per million tokens or a larger model at several dollars per million tokens.
For developers running local models on hardware they already own for other purposes (gaming, machine learning training, workstation tasks), the effective marginal cost of local inference approaches zero, which changes the break-even math considerably. The tool accounts for that distinction rather than assuming dedicated GPU purchases.
What This Means for Developers
The local versus API decision is rarely purely financial. Latency, maintenance overhead, model availability, and privacy all factor in. Sunk Cost does not pretend to make that entire decision for you. What it does is quantify the cost side with enough specificity that you can weigh it against the non-financial factors with actual numbers instead of gut feeling.
For teams running significant inference volumes, the calculator also exposes how sensitive the break-even point is to usage patterns. A developer processing a few thousand tokens per day will reach break-even much slower than one running continuous batch workloads. The tool makes those differences concrete rather than leaving them as abstract possibilities.