# AI agent cost calculator

Your agent re-derives the same workflow every run — and re-pays for it. This page turns a run schedule into a monthly number, and compares it to deterministic replay.

Interactive version: https://www.reelier.com/tools/agent-cost-calculator (inputs are encoded in the URL query — ?i=<5m|15m|1h|1d>&t=<tokens/run>&m=<haiku|sonnet|gpt|custom>&p=<custom $/MTok> — so any result is shareable as a link).

## The formula

No model behind this page — three multiplications:

1. `runs/month` — the interval over a 30-day month:

| Interval | Runs/month |
| --- | --- |
| Every 5 minutes | 8,640 |
| Every 15 minutes | 2,880 |
| Every hour | 720 |
| Every day | 30 |

2. `tokens/month = runs/month × tokens/run`
3. `$/month = tokens/month ÷ 1,000,000 × $/MTok`

Two deliberate simplifications, both of which UNDERSTATE the bill: every token is priced at the model's *input* rate (output tokens cost 5–6× more on every model listed, but tool-heavy agent runs are input-dominated — the benchmark run below was 18,390 in / 136 out), and retries, drift-fixing re-runs, and tool-result bloat are ignored. The real bill is higher.

## Defaults and where they come from

- **Tokens per run: 18,000** — measured average from the reelier benchmark (a real npm-registry agent workflow, N=10 on Claude Haiku 4.5: avg 18,390 input + 136 output tokens, ~$0.019/run). Methodology: https://github.com/seldonframe/reelier/tree/main/examples/benchmark
- **Model price presets** (input-token list prices, retrieved 2026-07-21):
  - Claude Haiku 4.5 — $1.00/MTok in ($5.00/MTok out) — https://platform.claude.com/docs/en/about-claude/pricing
  - Claude Sonnet 4.6 — $3.00/MTok in ($15.00/MTok out) — same source
  - GPT-5.6 Sol — $5.00/MTok in ($30.00/MTok out) — https://platform.openai.com/docs/pricing
  - Custom $/MTok field for anything else. Prices change — if these look stale, use the current pages.

## Worked example (the defaults)

Hourly × 18,000 tokens/run × $1.00/MTok (Haiku 4.5):

- runs/month = 720
- tokens/month = 720 × 18,000 = 12,960,000
- $/month = 12,960,000 ÷ 1,000,000 × $1.00 = **$12.96/month** — for one workflow, on the cheapest listed model, priced at input rate only.

The same schedule replayed with Reelier: **$0.00\*** — Level-0 replay makes no LLM calls, by construction. Verified from the run record on every replay (`llmInputTokens === 0`), not assumed.

\* The honest asterisk: replay re-executes your tool calls; API/egress costs of the tools themselves still apply. Judgment steps that genuinely need a model are opt-in escalation, not replay.

## What Reelier is

Record the agent run that worked, replay it deterministically — 0 tokens, byte-identical, a receipt on every step — and `reelier diff` fails the day it drifts.

Install: `npm i -g reelier && reelier init`

GitHub: https://github.com/seldonframe/reelier · npm: https://www.npmjs.com/package/reelier · Machine-readable site summary: https://reelier.com/llms.txt