How much does it cost to run one AI task?

It depends far more on how many tokens the task needs than on which model you pick. On published August 2026 rates, a single-page document extraction costs about 0.33 cents on Claude Haiku 4.5, while a 30,000-token contract review costs about 8 cents on Claude Sonnet 5 — roughly 24 times more on a comparable model. Batch processing halves both.

Does prompt caching actually reduce AI costs?

Yes, when a large part of your prompt repeats across calls. Anthropic prices cache reads at 0.1x the base input rate, so a stable system prompt, tool schema or document preamble costs a tenth as much after the first call. It does nothing for the parts of a prompt that change every request.

Is the cheapest model always the right choice?

No. At realistic volumes, inference is often a small part of total cost compared with build, integration and human review. A cheaper model that needs more human correction can cost more overall. Measure accuracy on your own data before optimising the token bill.

Tools

AI & LLM Cost Calculator

Work out what an AI workflow actually costs to run — per task and per month — using published API rates rather than per-token abstractions.

Step 1: What is the AI doing?

Each workload carries a typical token profile. Adjust your own numbers against these.

Step 2: Which model?

List prices per million tokens, read from each vendor's pricing page in August 2026.

Step 3: Volume and optimisations

Prompt cachingCache reads bill at 0.1x input. About 10% of this workload's input repeats.Batch processingThis workload answers a person in real time, so batch does not apply.

Cost per task

0.80 cents

Monthly at 10,000

$79.50

Annual

$954.05

13.5x the cost of GPT-4o mini on the same workload — a difference of $73.62 a month. Whether that is worth paying depends on accuracy on your data, not on the token bill.

Same workload, every model

Model	Per task	Per month	Per year
GPT-4o mini	0.06 cents	$5.88	$70.56
Claude Haiku 4.5	0.40 cents	$39.75	$477.02
Claude Sonnet 5selected	0.80 cents	$79.50	$954.05
GPT-4o	0.98 cents	$98.00	$1,176
Claude Opus 5	1.99 cents	$198.76	$2,385

How this is calculated

Cost per task is input tokens multiplied by the input rate, plus output tokens multiplied by the output rate. Prompt caching reprices the repeating share of input at 0.1x. Batch processing halves both sides. Nothing else is assumed.

Rates are vendor list prices read in August 2026. Token counts per workload are typical profiles, not measurements of your system — count the tokens on one real run of your own task and substitute them for a figure you can defend.

Deliberately excluded: retries, failed extractions needing a second pass, human review time, and the engineering cost of building the pipeline. In most deployments the human review line is larger than the inference line, which is why accuracy is worth measuring before cost.

What one AI workflow actually costs

The unit economics behind this calculator, worked through in full.

What RAG costs per 1,000 queries

Indexing, vector storage and generation priced separately.

Want this modelled against your real workload?

We build production AI systems and size the cost before the build starts, using your documents, your volumes and measured accuracy rather than list-price arithmetic.

See our AI & ML practice