Local vs. Cloud LLM Cost Calculator
Punch in your own token volume and see what routing part of it to a local model would cost versus paying cloud list prices for all of it. Nothing here is measured on your machine yet — it is an estimate from your inputs and public prices. The point is to show you where the break-even is, then go measure the real thing.
Your workload
Prices & hardware
sudo powermetrics on Apple Silicon.
Estimates from your inputs and public list prices — not measured until you run your own telemetry. Cheapskate's receipts are the measured version.
Example: the author's fleet
Over ~8 days of real routing (8,367 delegations), an always-on Mac served ~76% of tasks locally. Code-role work ran at a 2.5-second median latency with a 93% success rate; a domain-specific role ran 170 calls at 100%. That is where the 76% default above comes from.
Measured, not guessed — but measured on that hardware and that task mix. Your served-local share, throughput, and watts are what actually decide your number. Run your own telemetry.
How this is calculated
Transparency is the brand, so here is the whole thing. Cloud is priced straight off public per-million-token list prices. The local side is priced honestly as electricity for the active-inference time plus optional amortization — no hand-wavy "cost per token" that pretends to be measured.
Cloud cost (per month, 30 days):
Local active-inference time: the tokens you route locally, divided by your throughput, gives the seconds of active compute — which is what actually burns watts.
Local cost (per month): electricity for those hours, plus amortization.
Hybrid cost = local cost for the local share + cloud cost for the remaining share. Savings = cloud-only − hybrid.
What this deliberately does not do:
- It does not price a whole month of machine idle time — only the hours the model is
actively inferring (
local tokens ÷ tokens-per-second). Pricing 24/7 draw would be speculative. - It does not model retries or escalations. Cheapskate's real cost math includes retry and escalation rates; a public list-price calculator cannot, because those rates are specific to your models and tasks. That is exactly why the honest answer is "go measure it."
- It prices each call as a single one-shot request and ignores prompt-cache effects and multi-turn context growth, which would push cloud cost (and therefore savings) higher. So the savings figure is conservative.
- Amortization is the only fixed monthly cost, so the break-even volume is the daily token volume at which the electricity + amortization on the local share equals what that same share would have cost in the cloud.
Built on Cheapskate — the routing layer that sends each task to the cheapest model that passes your quality bar, local or cloud, and shows you the receipts.
AI Product Architect & Strategic Advisor · 24 years of applied ML · Co-inventor of Real-Time Bidding (RTB)