AI Compute RadarCost check
What does it cost?
Three honest ways to run the same model: pay per token for the full-precision API, or run its quantized weights on a machine you own or rent by the hour. Pick your usage and hardware — every number below decomposes into a measured price, a sourced vendor figure, or an assumption you can change.
Cheapest at your settings: Cloud API
Money only — the paths differ in speed, privacy and, for quantized weights, output quality.Cloud API — Cheapest at your settings
per day
$0.334
per week
$2.34
per month
$10.15
per year
$122
$0.150 in / $1.88 out per 1M tokens
Your machine
◇ Estimated— Estimated: computed from our curated model and hardware catalog — not a live reading.per day
$2.01
per week
$14.07
per month
$61.15
per year
$734
$1.46 hardware + $0.550 electricity per day — runs the Q4_K_M quantized weights
Rented GPU
per day
$1.07
per week
$7.49
per month
$32.53
per year
$390
RTX 3090 at $0.267/h — cheapest measured class that fits the Q4_K_M weights
Every assumption on the table
- + Usage: 500,000 tokens per day, of which 70% input tokens — an order-of-magnitude anchor, not a measurement.
- + Active compute: 4 hours per day, used for electricity and rental time alike.
- + Power draw: 450 W board power (vendor spec) plus 100 W for the rest of the machine, while active.
- + Hardware spread over 3 years, straight-line; resale value ignored.
- + All figures are USD. API and rental prices are the same worldwide; hardware list prices are US figures excluding sales tax (EU retail prices include VAT and differ), and electricity varies a lot by country — edit both fields to your local numbers.
- + A month is 365/12 ≈ 30.4 days, so twelve months equal one year.
- + API and rental prices are collected automatically and carry their collection time; hardware prices are vendor list prices (launch MSRP or current official price, as noted next to the field) you can override.
- + The three paths differ in speed, privacy and — for quantized local weights — output quality. This table compares money only.
Reading the cost calculator
As of today's readingsOn November 2, 2026 the calculator offers 27 of the 49 tracked models: those with a measured API price and a measured local file. Among them the API input price runs from $0.01 per million tokens (GLM-5.2) to $2.00 (Qwen3.8 2.4T-A95B). The rental column stands on today's Vast.ai medians, among them RTX 4090 $0.51, H100 $3.08 per GPU-hour. The own-hardware column can be set to any of 37 tracked machines.
The calculator answers one question in three columns: what the same model costs per day when you pay an API per token, when you run its quantized weights on a machine you own, and when you rent a GPU by the hour. The three numbers are built from different kinds of input, and the page keeps them apart on purpose: the API column uses a measured list price per million tokens, the own-hardware column a sourced launch price spread over a declared number of years plus a declared electricity price, the rental column the marketplace median of the cheapest class that runs the model comfortably.
Every assumption is yours to change: tokens per day, active hours, the hardware you own, its purchase price, the years you spread it over, the price of a kilowatt-hour. The defaults are not recommendations but starting points, and the page names each one next to its number. Two things are left out and stay out: speed and quality. A rented H100 and a desktop card do not produce tokens at the same rate, and a quantized file does not answer exactly like the full-precision API; this page prices capability, not performance, and says so in its own disclaimer.
Only models with both a measured API price and a measured local file appear in the list, because a column without a measured number would have to be invented. When a model disappears from the list it is usually because its API price vanished from the catalogue, not because it stopped running locally; the model page still shows its memory table.
What this is not
Not a quote and not advice: prices move, electricity varies, and your real token volume is yours to judge. Each column shows where its numbers come from — check them against the live sources before spending money.