Multi-Model Agent Cost Calculator
Compare what one agent costs to run across several models or providers — per run, per day and per month, side by side. You enter the prices (input, output and optional cached-input $/1M tokens) and describe the workload once — tokens in and out per step, steps per run, runs per day, and an optional recursion/meta multiplier — and the tool costs every model against it, flags the cheapest and most expensive, charts the monthly spread, and shows where the money goes between input and output. The calculator ships no vendor price list, so its numbers can never be stale — they are only ever as current as what you type. Everything runs in your browser; nothing is uploaded, and the math is a labelled heuristic estimate, not a billing quote.
Rows marked “example — edit me” carry placeholder prices only — replace every price with your own current rates. This tool bakes in no vendor price list, so its numbers are only ever as accurate as what you type.
Models & prices
3 / 24Cost comparison
estimatePer-run, daily and monthly cost for each model under the workload on the right. Month = 30 days.
| Model | Per run | Daily | Monthly |
|---|---|---|---|
Frontier model (example — edit me) priciest | $0.187 | $37.44 | $1,123 |
Mid-tier model (example — edit me) | $0.062 | $12.48 | $374.40 |
Small / fast model (example — edit me) cheapest | $0.021 | $4.20 | $126.00 |
Frontier model (example — edit me) costs $997.20/mo more than Small / fast model (example — edit me) — that is 8.9× the cost for the same workload.
Monthly cost
Where the money goes
Input (prompt, incl. cache) vs output (completion) share of each model's run cost.
Workload
Steps are the tool calls / reasoning turns per run. The meta multiplier models recursion / self-improvement overhead — set it from the Recursion Budget Simulator. Month is treated as 30 days.
The formula (no magic)
cachedIn = tokensIn × cachedPct/100
freshIn = tokensIn − cachedIn
stepIn = freshIn/1e6 × inPrice
+ cachedIn/1e6 × (cache ?? inPrice)
stepOut = tokensOut/1e6 × outPrice
perRun = (stepIn + stepOut)
× steps × meta
daily = perRun × runsPerDay
monthly = daily × 30A heuristic estimate, not a quote. Real spend varies with caching behaviour, batching, rate limits and per-provider billing rounding. No prices are baked in — everything above comes from what you entered, so nothing here can go stale.
What this tool is
- A free, fully in-browser calculator for comparing what one agent costs to run across several models or providers — side by side, per run, per day and per month. Nothing is uploaded; there are no server calls and no account.
- It ships NO vendor price list. You enter every price yourself — input $/1M tokens, output $/1M tokens and an optional cached-input price — so the numbers can never be stale or wrong; they are only ever as current as what you type.
- You describe the workload once — tokens in/out per step, steps (tool calls / reasoning turns) per run, runs per day, and an optional recursion/meta multiplier — and every model is costed against it.
- Outputs: a side-by-side table with cheapest vs most-expensive and the monthly delta, an inline-SVG bar chart, and an input-vs-output "where the money goes" split. Export to Markdown or CSV, or copy a shareable permalink that encodes the whole comparison in the URL.
- The math is shown in full and labelled a heuristic estimate — a planning aid, not a billing quote. Real spend varies with caching, batching, rate limits and per-provider rounding.
Cost is an architecture decision, made before you ship
The cheapest model is rarely the cheapest agent. A frontier model that needs one pass can beat a cheap model that needs five retries, and a heavy prompt cache can turn an expensive input price into a rounding error. The only way to see which model actually wins for your workload is to put the per-token prices, the token volumes, the step count and the recursion overhead in one place and let the arithmetic settle it. That is all this tool does — and because it holds no prices of its own, it is equally valid for a public API rate, a private deployment, or a rate you negotiated.
The deliberate omission here is a price list. A tool that hard-codes vendor prices is wrong the day a vendor changes them, and worse, it is confidently wrong — it still produces a clean-looking number. By making the prices your input, the calculator trades a little typing for the guarantee that it can never quietly go out of date. Pair it with the Recursion Budget Simulator to derive the meta multiplier for a self-improving or fan-out loop, and with the Custom Agent Harness Designer to see how harness choices — parallel branches, verifier passes, cost-aware model routing — change the step count and the multiplier in the first place.
Frequently asked questions
Why does this tool not include current model prices?
On purpose. Model and provider prices change often, and a baked-in price list is wrong the moment a vendor updates it — which quietly makes every calculation wrong too. By asking you to enter the prices yourself, the tool can never be stale: the numbers are exactly as current as the rates you paste in from the provider you are actually comparing. It also means the tool works for any model, provider, private deployment or negotiated rate, not just the ones someone remembered to hard-code.
How is the per-run and monthly cost calculated?
For each model: cached input tokens = tokens-in × cached %, and fresh input tokens are the rest. Step input cost = fresh/1e6 × input price + cached/1e6 × (cached price, or the input price if you left it blank). Step output cost = tokens-out/1e6 × output price. Per-run cost = (step input + step output) × steps × meta multiplier. Daily = per-run × runs/day, and monthly = daily × 30 days. The full formula is printed on the page, so nothing is hidden.
What is the recursion / meta multiplier for?
Agents that recurse — self-improvement loops, parallel Propose·Probe·Compose fan-outs, repeated verification passes — do more than one model pass per logical run, which multiplies token spend. The meta multiplier is where you fold that overhead in: a 3× multiplier triples the per-run cost. If you are not sure what number to use, the Recursion Budget Simulator estimates it for a given loop shape, and you carry that figure here.
Are these numbers a billing quote?
No. They are a transparent, deterministic estimate computed from the prices and workload you entered — a planning and comparison aid. Actual invoices vary with prompt-cache hit rates, request batching, rate-limit retries, minimum billing increments and per-provider rounding. Treat the output as "roughly how these options compare under my assumptions", not as a figure to put in a contract.
Does anything leave my browser?
No. The whole tool is client-side React with inline SVG. There are no network or API calls, no account and no storage of your inputs on any server. The shareable link encodes your models and workload directly in the URL hash (base64url), so sharing it is your choice and the data travels only in the link you copy. Markdown and CSV export both run locally.