What an AI agent costs to run
An agent is not one call. It takes a number of steps, and the context it reads grows with every one of them. Set the shape of your workload, pick a model, and the calculator prices the whole run at the maker’s published rate, step by step.
Free and without sign-up. The arithmetic runs in your browser, so nothing you type is sent anywhere.
$2 in and $10 out per million tokens, the price Anthropic publishes, read on 22 September 2026.
One model call each: a search, a read, a decision, a write. Capped at 40.
The instructions, the tools it is given and the request itself. Roughly 750 words per 1,000 tokens.
What the step brings back: the page it read, the rows it fetched, the result of the tool it called.
The answer only. Reasoning is the next field, because it is billed at the same rate.
Billed at the output rate. Leave it empty for 400, an assumption; set it to 0 for a model that does not think first.
Leave it empty to use 70%, an assumption. The first step is a cold read, so none of it is cached.
Leave it empty to use 5%, an assumption.
How often the whole workflow runs: per ticket, per lead, per document.
- Read over 10 steps
- 107,500 tokens
- Of that, served from cache (70%)
- 72,450 tokens
- Written: the answer and the reasoning
- 9,000 tokens
- One run, 5% of steps retried
- $0.1833
- 1,000 runs a month
- $183
List price as Anthropic publishes it, read on 22 September 2026. Cache hits at $0.20 per million. Hosting, the tools the agent calls, embeddings and logging are not in it.
$183. What 1,000 runs of a 10-step agent cost a month on Claude Sonnet 5.
| Step | Reads | From cache | Writes | Cost |
|---|---|---|---|---|
| 1 | 4,000 | 0 | 900 | $0.0170 |
| 2 | 5,500 | 3,850 | 900 | $0.0131 |
| 3 | 7,000 | 4,900 | 900 | $0.0142 |
| 4 | 8,500 | 5,950 | 900 | $0.0153 |
| 5 | 10,000 | 7,000 | 900 | $0.0164 |
| 6 | 11,500 | 8,050 | 900 | $0.0175 |
| 7 | 13,000 | 9,100 | 900 | $0.0186 |
| 8 | 14,500 | 10,150 | 900 | $0.0197 |
| 9 | 16,000 | 11,200 | 900 | $0.0208 |
| 10 | 17,500 | 12,250 | 900 | $0.0219 |
Ten steps, a context that grows from 4,000 tokens, 70% of each read served from cache and 5% of steps retried: on Claude Sonnet 5 that run reads 107,500 tokens and costs $0.1833, so 1,000 runs a month cost $183. Its published batch rate halves that.
How the number is worked out
Every figure the calculator assumes is on this page, so the result can be checked with a pencil.
- The prices
- Input, output and cache-hit rates as each maker publishes them, read from data/models/2026-09/models.json, which records the pricing page and the day it was read (22 September 2026). The same figures are on our model pages. No price here comes from a reseller.
- The steps
- Step one reads the starting context. Every step after it reads what the step before read plus the tokens that step added, so the input billed over a run is the sum of the steps, not one prompt. The per-step table prints each line.
- The cache
- The share you set is served at the maker’s cache-hit rate; the first step is a cold read, so nothing of it is cached. Where a maker publishes no cache price, cached tokens are billed at the full input rate and the result says so.
- Reasoning and retries
- Reasoning tokens are billed at the output rate, so they are added to the answer for every step. The retry rate is the share of steps that fail and run again: it multiplies the run by one plus the rate. Both defaults are assumptions.
- The batch rate
- Where the maker publishes a batch discount in its price notes, the calculator offers it and prints the share: half off on Anthropic, OpenAI and Google, a fifth off on Grok 4.3, and nothing where the maker does not sell one.
- What it leaves out
- Hosting, tools the agent calls, embeddings, logging, and the long-context surcharges that start above 200,000 or 272,000 tokens in a single request on some models. Keep a step under that band and the base rate is the one that applies.
Questions people ask about this number
Have the run priced on your own workflow
The audit measures the real steps, the real context and the real volume before anything is built, so the number you plan with is measured rather than assumed.