How often is the pricing data updated?

The catalog refreshes hourly. Each row in the table shows the exact "last verified" timestamp alongside a link to the provider source it was read from, so you can confirm a price yourself rather than trusting a stale snapshot.

Do I need an account to compare prices?

No. Browsing, filtering, reading the chart, copying share links, and using the ROI calculator are all account-free.

What does "quality tier" mean on the chart and table?

It's a rough capability class — frontier, balanced, or budget — used to group models with similar strengths, not a numeric benchmark score. Use it to avoid comparing a narrow classifier against a frontier reasoning model as if they were substitutes.

Why do some models show no current price?

When a provider changes a rate by more than a review threshold, the new price is held for manual confirmation instead of being published automatically, so a model can briefly show no current price rather than an unverified one.

Is the price shown here what I'll actually be billed?

Treat it as an informational estimate, not an invoice. Provider pricing pages are the source of truth — this table exists to help you compare quickly, not to replace checking directly before you commit spend.

Workload planningAll guides

How to estimate monthly token usage

Last updated 2026-07-18

Abstract hand-drawn monthly token usage calendar with input and output buckets — Original AICostCompass illustration. Abstract generated artwork; no people or animals.

You do not need perfect telemetry to create a useful AI budget. You need a reasonable model of monthly volume. The best starting point is simple: count how often the workflow runs, estimate how much text goes into the model, and estimate how much text comes back. Keep input and output separate, because providers often price them differently.

Token budgeting workflow

Step 1

Count monthly events: tickets, chats, pull requests, documents, calls, or generated assets.

Step 2

Estimate input per event: instructions, history, retrieved context, files, and user text.

Step 3

Estimate output per event: answers, code, summaries, JSON, tool calls, or drafts.

Start with the business event

Use the real event that triggers the model call. A support bot may run per conversation or per message. A code review assistant may run per pull request. A document workflow may run per uploaded file. A meeting assistant may run per transcript. Once you can count the event, monthly cost becomes easier to forecast.

Estimate input honestly

Input includes more than the user prompt. Count the system prompt, examples, retrieved documents, tool results, conversation history, schemas, and file excerpts. Many teams underestimate this because the visible user message is short while the hidden context is large. If a workflow uses retrieval, estimate how many chunks are usually inserted.

For rough planning, create three scenarios: low, expected, and high. The high case matters because successful AI features often grow in usage. A pilot with 1,000 monthly requests can become 100,000 quickly if the workflow actually helps people.

Estimate output separately

Output is where many estimates break. A classifier may return 50 tokens. A report writer may return 3,000. A coding assistant may produce code, comments, tests, and explanation. Since output is often more expensive, a small change in answer length can shift monthly cost and provider ranking.

Convert to monthly totals

Multiply events by average input and output. For example: 20,000 support conversations per month, 1,500 input tokens each, and 250 output tokens each equals 30 million input tokens and 5 million output tokens. Put those numbers into the comparison tool rather than trying to eyeball unit prices.

References and fact checks

Anthropic Claude API pricing - shows separate input, output, cache write, and cache hit units
Google Gemini API pricing - documents input/output tokens plus media and batch-related billing concepts
AWS Bedrock cost report documentation - useful for reconciling estimated token usage with billing data

How to do this in AICC

Turn the article into an answer you can use

Use AICC to turn rough workload assumptions into monthly input and output volumes, then check how those volumes affect provider cost.

AICC workflow map

Move 1

Count monthly events

Move 2

Estimate input and output

Move 3

Compare provider cost

1
Write down the triggering event
Before opening the tools, define the event: support conversation, document, pull request, meeting transcript, generated report, or chat turn. This is the unit you will multiply by monthly volume.
2
Estimate input and output separately
Use the article method to estimate average input tokens and output tokens per event. Keep prompt, retrieved context, chat history, and generated answer length separate.
3
Enter the monthly totals in Compare
Open the Comparison tool and enter total monthly input and output volumes. If the workflow also uses images, video, audio, seats, or volume assumptions, fill those fields too.
4
Stress-test the high case
Increase the monthly volume or output length to see where cost changes materially. This teaches you which assumption matters most before the invoice arrives.

What you should have at the end

You should have low, expected, and high monthly usage cases and a provider comparison that reflects your actual workload shape.

Comparison tool

Enter your workload assumptions and compare estimated monthly cost across selected tools.

Pricing map

Scan providers, model tiers, unit prices, quality signals, and visible market options.

ROI calculator

Translate AI spend into saved time, labor value, and business return.

Frequently asked questions

What's the biggest mistake teams make when estimating input tokens?

Undercounting hidden context — the visible user message is often short, but system prompts, retrieved documents, tool results, and chat history can make real input volume much larger.

Why build low, expected, and high usage scenarios instead of one estimate?

Successful AI features tend to grow quickly in usage — a pilot with 1,000 monthly requests can become 100,000 once it proves useful, and the high case protects your budget plan.

Should I estimate input and output tokens together or separately?

Separately. Providers typically price them differently, and since output is often the pricier side, small changes in answer length can shift which provider is cheapest.