AI API costs · USD

Estimated AI Usage Costs

ESTIMATED TOTAL COST PER MONTH
Expected
Cost / user / month
Expected
Total monthly API cost
Expected
Requests / month—
Cost / request—
AI model:
What this means

What to do

Cost anatomy

Where the bill comes from

Live token sensitivity

How token size changes cost per user

User prompt AI reply

COST PER USER / MONTH

— Expected

Each line changes only that token size while keeping the rest of your Expected scenario fixed. The Y-axis shows API cost per active user per month.

AI API cost per user per month by prompt or reply token size Orange shows prompt-token sensitivity in API cost per user per month. Dark orange shows reply-token sensitivity. Expected settings are marked on each line and Max Expected token settings are shown in magenta when range is enabled. User prompt AI reply COST / USER / MONTH TOKENS PER PROMPT / REPLY
Prompt expected — Reply expected —
What this means

What to do

Request size guide

What does “500 tokens” actually look like?

Rough English-text examples. They are orientation, not a tokenizer. Your current settings are highlighted and update with the controls.

User prompt

Everything sent to the model: instructions, question, pasted text, documents and chat history.

Your setting
—
Short message · ~75 words
~100
Detailed task · ~375 words
~500
Several pages · ~3k words
~4k
Large document/context
~20k

AI reply

The billable model output. For reasoning models, provider-reported reasoning tokens can also matter.

Your setting
—
Short answer · ~110 words
~150
Detailed answer · ~500 words
~700
Long article / code · ~1.5k words
~2k
Large code / document
~8k
Deep research, images, audio and video: do not force these into one fake “token equivalent”. Deep-research/agent workflows often make many model and tool calls, so increase AI requests per user as well as prompt/reply size. Image, audio and video billing varies by provider and model and may use separate units or prices; this estimator intentionally calculates standard text-token API usage only.
What this means

Text size is easy to estimate roughly; agentic and multimodal work is not. A single visible user action can create several billable calls behind the scenes.

What to do

Prototype one real task, inspect the provider usage/telemetry, then enter the measured calls and token counts here. For generated images or video, budget that modality separately.

Decision order

Three cost levers to test first.

There is no honest universal percentage split: the weight depends on your model and workload. These are the levers with the clearest effect.

1 · Model

Pass the quality test first

Same workload, different unit price.

How to find out: run the same 10–20 real tasks on a few models and compare quality, latency and cost.

2 · Request frequency

Count what one user action really triggers

Within the same model and token sizes, 2× as many API calls means about 2× the API cost.

How to find out: log actual API calls per user action, especially for agents, retries and tool loops.

3 · Prompt + reply size

Measure payload, history and output

Within the same model, doubling both input and output tokens roughly doubles token cost.

How to find out: inspect real usage tokens and watch repeated history, documents and long reasoning/output.

Guardrails

Keep usage under control

Uncapped burst cost

What no rate limit can cost you

Someone scripts a public form and fires requests in a loop. No login, no rate limit, no daily cap — it happens to free tools constantly. This stress test shows what that burst costs at your current prompt size.

Current burst stress test: —.

What this means

A public AI feature without limits is a blank cheque. Max expected is a planning scenario, not a spending limit.

What to do · high priority

Before public launch, set per-user rate/daily limits, provider spend alerts or caps where available, and a sane retry/tool-loop ceiling.

Use it well

Get a useful range, not one magic number.

Who it is for

Founders, product owners, indie devs, agencies and technical PMs planning or pricing an AI feature.

Best way to use it

Start broad. Then keep everything fixed and change one thing at a time. Compare a few models and scenarios.

What it is not

Precise production billing. If you already have real usage telemetry, use that instead.

Simple planning rule: use Expected for your normal forecast and pricing maths. Use Max expected as the affordability check: the product should still be safe if that month happens. If Max expected would hurt, add limits, change the model, reduce usage, or change pricing before launch.
Model comparison

Same workload. Different model.

All bars use the same cost scale. Select a model to update the estimate; test quality on your own tasks.

Monthly API costs for the same workload, sorted by expected cost
Model Expected

Alternatives are the lowest-cost models in the supplied frontier and fast tiers. These tiers are editorial guidance, not a guarantee of task quality.

What this means

The same workload can cost wildly different amounts across models. Model choice is usually the first lever to test before squeezing prompts or betting on caching.

What to do

Shortlist two or three models, run the same real tasks through each, and choose the cheapest one that still clears your quality and latency bar.

Model fit

Model fit

Consider it for

Think twice for

Fit notes are supplied editorial guidance. Validate with your own examples.

Details
Detailed assumptions & pricing

Calculation: users × requests per user/day × active days × token cost per request. Input cost includes the cache-hit share at the cached rate; output is billed separately.

Standard text-token rates only. Excludes cache-write and storage fees, tools, images, audio, retries, batch discounts and taxes. Include all billable output in your output-token estimate. Prices and model availability are not fetched live.

Plain English

What these words mean.

TokenA chunk of text, roughly ¾ of an English word. 1,000 tokens is roughly 750 words, but language and content change the ratio.

Input / user promptEverything you send: the question, system prompt, attached text/documents and relevant chat history.

Output / AI replyWhat the model writes back. Reasoning models may also report billable reasoning/output tokens depending on provider.

Cached inputRepeated context the provider can serve at a lower rate when its cache rules are actually met.

Context windowThe maximum the model can hold at once. It is a limit, not the amount you automatically pay for.

Batch / Flex / FastAlternative processing modes some providers offer. This estimator uses its saved standard-rate dataset unless explicitly stated otherwise.

Rate limitA cap on how often a user or client can call the API. One of the cheapest protections against runaway usage.

Edit estimate

Changes update the estimate immediately.

Copy your report

Automatic copying is unavailable here. Select and copy the text below.