How it works

Token

The unit AI providers use to measure text. Roughly ¾ of a word in English — "ModelPricing" is 3 tokens, a 1,000-word article is roughly 1,300 tokens. Providers charge separately for input tokens (text sent to the model) and output tokens (text the model generates). Output tokens cost significantly more because they require more computation.

Workflow

One complete task a user triggers — e.g. "analyse this contract", "book an appointment", "qualify this lead". A workflow might involve multiple model calls, tool lookups, and document reads. It is the natural unit for measuring usage volume: how many tasks does your system run per month?

Model call

A single API request to the language model. One workflow often requires several calls — for example: extract data → reason about it → generate a reply → check the output. Each call consumes input and output tokens and is billed separately. Multi-step agents typically have 2–5 calls per workflow.

Retry overhead

Not all calls succeed first time. Some fail due to rate limits, JSON parse errors, or model refusals — triggering an automatic retry. Retry overhead % adds a fraction of extra calls on top of your expected call count. A 5% retry rate means 5 extra calls per 100 attempts.

Cache hit rate %

When the same system prompt or context appears in repeated calls, providers can cache the input and charge a fraction of the normal input rate (typically 10–20× cheaper). The cache hit rate is the percentage of your input tokens that get served from cache rather than reprocessed. A customer support agent with a fixed knowledge base might see 30–50% cache hits; a general assistant with unique per-user context might see 5–15%.

Batch share %

Providers offer a batch API for non-real-time work — jobs submitted in bulk and completed within hours rather than seconds. Batch pricing typically gives a 50% discount. The batch share is the percentage of your total calls that can tolerate this delay. Report generation, nightly summarisation, and data enrichment pipelines are good candidates; live chat is not.

Batch discount rate %

The discount applied to batch calls. Most providers offer 50%. This is pre-filled from the model preset but editable if your agreement differs. The discount applies to the entire call cost (input + output), not just one tier.

Web searches per workflow

The average number of search API calls made per workflow. A simple chatbot that sometimes looks things up might average 0.3. A research agent that always searches might average 3–5. Set to 0 if your system does not use web search.

Search content tokens

After a search call returns results, those results are pasted into the model's context as additional input tokens. Search content tokens is the average number of tokens retrieved per search call. Search results are typically 1,000–4,000 tokens per call. These are billed at your model's input rate, because they enter the model as input.

Embedding

A numerical representation of text — a list of numbers (a "vector") that captures the meaning of a piece of text. Embeddings are used to power semantic search and RAG (retrieval-augmented generation) — finding relevant documents from a knowledge base based on meaning, not keyword matching. Embedding models are separate from language models and typically much cheaper.

Monthly embedding tokens

The total number of tokens you embed each month — both for indexing your knowledge base and for embedding user queries at runtime. A 10,000-document knowledge base with 500 tokens per doc = 5M tokens to index (done once). Embedding each user query might add another 100 tokens × number of queries per month.

OCR — Optical Character Recognition

Converting images or scanned PDFs into machine-readable text. AI document pipelines (legal, finance, healthcare) often need to extract text from uploaded PDFs before the language model can read them. OCR APIs (e.g. AWS Textract, Google Document AI) charge per page processed. Set to 0 if your system only handles plain text or already-digital documents.

OCR pages per workflow

The average number of document pages processed in a single workflow. A contract review agent might process 8–15 pages per run. A simple form processor might handle 1–2 pages. The total monthly OCR pages = workflows × pages per workflow.

Fixed infrastructure cost

Your monthly server, hosting, and platform costs — regardless of traffic. This includes: compute (EC2, Cloud Run, Fly.io), API gateway, monitoring (Datadog, Sentry), CI/CD pipelines, and any third-party orchestration services (LangSmith, Helicone). Does not scale with usage.

Vector storage cost

If your system uses a vector database for semantic search (Pinecone, Weaviate, Qdrant, pgvector), it has a monthly storage cost. This is typically flat-rate at low scale and scales with the size of your index. Included as a separate field because it is often the second-largest infrastructure line item in RAG systems.

Contingency buffer %

A safety margin applied to the entire cost estimate. Real usage rarely matches projections exactly — token counts vary, traffic spikes, caching behaves differently than expected. A 10–15% contingency is standard for planning purposes. It does not represent a real charge — it is a budget cushion.

API cost / month

What you would owe AI providers at P50 volume. Calculated from published rates × your volume estimates. The rates are sourced from provider pricing pages; the volumes (tokens per call, retry %, cache-hit %) are your estimates and carry the same uncertainty as operating overhead inputs.

API cost per workflow

API cost divided by total monthly workflows. The most useful unit for product decisions — tells you what one task costs in AI charges before overhead. Use this to compare models or price individual features.

Operating overhead / month

Your fixed infrastructure costs (infra + vector storage). Based on your own estimates — the calculator does not compute these. Does not include contingency, which is shown separately in the budget view.

Total estimate / month

Planning total: API cost + operating overhead + contingency buffer. Used for budget approval. Profit, gross margin, and break-even use the P&L total (without contingency) — because contingency is a planning reserve, not a realized expense.

Cost per active user

(API cost + operating overhead) ÷ monthly active users. Excludes contingency. This is your unit economics baseline — the minimum revenue per user needed for the product to cover its direct costs.

Price for target gross margin

Uses only variable API costs per user (the correct COGS for gross margin). Formula: variable_cost_per_user / (1 − target_GM). This matches how gross margin is defined in a standard P&L — revenue minus direct cost of revenue, divided by revenue. It excludes fixed infra, monitoring, and contingency, which belong below the gross margin line. For a full operating-margin price, add fixed cost allocation on top.

Break-even users (CVP)

Uses the standard cost-volume-profit formula: fixed costs ÷ (revenue per user − variable cost per user). This is the volume at which contribution margin covers fixed costs. If variable cost per user exceeds revenue per user, there is no break-even — the calculator surfaces this as a warning. Note: excludes contingency from both fixed and variable costs.

Gross margin %

Only shown when you enter a revenue per user. (revenue − API costs) / revenue × 100. This is true gross margin — revenue minus variable COGS only. Fixed infra and contingency sit below the gross margin line. A SaaS business typically targets 60–80% gross margin on this basis.

Yearly run-rate

Planning total × 12. This is a flat run-rate, not a projection. It does not account for provider price changes (historically 20–40% annual decline), user growth, or discounting. Use it as a rough annual budget ceiling, not a financial forecast.