Estimate the AI API cost of your SaaS.
A builder subscription helps you make the product. Once customers use it, each model call can add to a separate API bill. Count that variable cost before choosing a paid plan price.
One usage pattern, five model prices
Suppose 100 people are active in a month. Each makes 30 model calls, with 2,000 input tokens and 500 billed output tokens per call. That is 3,000 calls per month. The same workload costs:
| Model | Input / 1M | Output / 1M | Monthly API cost |
|---|---|---|---|
| Muse Ember | $0.10 | $0.50 | $1.35 |
| GPT-6.1 Terra | $1.00 | $5.00 | $13.50 |
| Gemini 4 Hydrogen | $1.50 | $7.50 | $20.25 |
| Claude Sonnet 5.5 | $2.00 | $10.00 | $27.00 |
| Claude Fable 5.2 | $8.00 | $40.00 | $108.00 |
USD, standard rates verified 2026-11-13. This is an illustrative, text-only estimate with no cache hits, retries, tool fees or long-context premiums. It is not a prediction of what users will do. The model catalogue supplies the rates; change the assumptions in Cost Lab before making a pricing decision.
Start with calls per customer
List each action a customer can take: ask a question, summarize a document, generate a report. Record how many model calls one action triggers. An agent may make several calls before returning one answer, so a count of button presses will understate its cost.
For each call, measure the full input: system instructions, user content, retrieved passages and tool results. Then record billed output from the provider's usage report, including reasoning tokens where the provider charges for them. Paste representative inputs into the token calculator to plan the input side.
The basic monthly formula is active users × calls per user × (input tokens × input rate + output tokens × output rate) ÷ 1,000,000. Use a busy customer as well as an average one to see where a generous usage allowance becomes expensive.
Recheck the bill when the workflow changes
A larger system prompt, a document attached to every request, longer reasoning output or extra agent steps can move the bill even if user count stays flat. Log model calls and tokens per completed task, then compare that number with the estimate you used to price the plan.
Repeated prompt prefixes may be cheaper through caching, but only when they actually hit the cache. Compare cached and uncached agent costs. If you send a whole document on every request, compare long context with retrieval before scaling.
Include the rest of the launch budget
API tokens are only one line. Add the builder or coding tool, hosting, database, email, monitoring, payment fees and the time needed to support customers. Some builder plans bundle hosting or model usage; check the allowance and what happens when it runs out.
For the wider path from validating an idea to an MVP, payments and first customers, read Jobbit's launch guide:
Read on Jobbit.ukHow to Build a SaaS With AI in 2026: From Idea to Paying CustomersIts launch checklist complements the per-customer API budget above.
Measure against the model you use
Other comparisons worth running
Compare approaches, not just models
Frequently asked questions
How do I estimate an AI SaaS API bill?+
Multiply active users by model calls per user, then by the input and output token cost of one call. Count every model call in an agent workflow, including retries and tool follow-ups.
Does this example include hosting or a builder subscription?+
No. The table includes only uncached text input and output at standard API rates. Hosting, storage, email, payment fees, builder subscriptions and tool charges need separate budget lines.
Why can the real bill differ from the table?+
Actual prompts, output lengths, cache hits, reasoning tokens, long-context tiers and service tiers change the bill. Measure provider-reported usage on representative tasks before setting a paid plan price.