What does your prompt cost?
Every prompt has a price. Know it before you send.
Paste anything. See token counts and price across every major model — computed in your browser, not on a server.
Gateways and monitoring tools optimize spending you already have. Token Bench tells you what you’ll spend before you commit.
Monthly cost by model
Reply length is converted to tokens separately for each model, using how that model tokenized your text. Reasoning models bill their hidden thinking as output; if you use one, enter a typical figure from your API usage logs. The repeating part is anything sent unchanged every time, such as a system prompt, and is billed at the cached rate.
| Model | In /1M | Out /1M | Context | Per request | Per month | vs cheapest |
|---|
A month is 30 days. Rows marked est. use estimated token counts; use “get exact count” above for Claude and Gemini. Cache-write charges, batch discounts and negotiated rates aren’t included.
Mix two models
route routine requests to a cheaper modelThe success rates are example values: replace them with what you measure by testing each model on your own requests. Costs are per completed task. When the routine model fails, the request is sent on to the hard model; when the hard model fails, it’s retried. Uses your prompt and the usage settings above, and assumes routine and hard requests are about the same size. It doesn’t include the cost of deciding where each request goes.
Batch & repo estimator
Drop many files at once for a combined token and cost total. Files are read locally and never uploaded.
| File | Chars | OpenAI | Claude | Gemini |
|---|---|---|---|---|
| No files yet | ||||
Open-model tokenizers
Exact counts for Llama 3, Mistral and Grok 1. The real tokenizer downloads once from Hugging Face, then runs offline in your browser.
Prompt compressor
Strips filler, hedge words and citations to cut tokens. Rule-based, so dense technical prose may compress very little.
Multimodal calculator
Image, video and audio token counts using Google's published Gemini formulas.
Caching savings
Model the cost of sending the same prompt repeatedly with prompt caching enabled. Uses the text from the main calculator.
Token heatmap
See which parts of a prompt cost the most. Darker shading means more tokens for each character.
Where your tokens go
| Content | Share of text | Share of tokens | What to do |
|---|
Most expensive paragraphs
top 6 by tokens| Paragraph | Characters | Tokens | Share | Chars per token |
|---|
Shading uses OpenAI’s o200k tokenizer, the same one behind the calculator’s OpenAI count. Other models split text differently, so exact numbers vary, but the same kinds of text tend to be expensive everywhere. Point at or tap any piece, or select a passage, to see exactly what it costs.
Token Game
Which text costs more AI tokens? Guess right to keep your streak going, and learn why as you play.
Counted with o200k, the tokenizer behind OpenAI’s recent models. Other models split text differently, so their counts would differ, but the same kinds of text tend to be expensive everywhere.
Prompt Builder
Build a prompt from blocks. Every brick is one token. Fit the job inside the budget, without leaving out anything the job needs.
Your prompt
Your prompt in cubes
One cube per token. Cubes above the dashed budget line turn red.
Blocks
Bricks are counted with o200k, the tokenizer behind OpenAI’s recent models. Savings are worked out at 10,000 requests a day on GPT-6.1 Sol, where every token in a prompt costs about $0.60 a month. In Hard, the reply’s length is an assumption based on typical replies: about 50 tokens for two sentences, 80 for 60 words, and 260 when nothing limits it, with Hindi replies about 2.6 times longer in tokens. Real replies vary.
Token Academy
Learn what AI tokens are, why they cost what they do, and how to spend fewer, through short games. Three levels, about fifteen minutes. No sign-up: your progress is saved in this browser.
Research
Original measurements of what AI tokens cost, and what new research means for your bill.
Browser extension
See cost while you type — not after.
Live cost inside ChatGPT, Claude and Gemini
A small overlay shows your running token count and cost as you write, so you catch an expensive prompt before you send it. Counting happens locally — no text leaves the page.
Coming soonSubmitted to the Chrome Web Store and awaiting review. This page will link to the listing once it’s approved.
API
Token counts and cost for any text, across every supported model. Free to use while in beta.
curl -X POST https://whatdoesmypromptcost.com/api/v1/tokenize \
-H 'content-type: application/json' \
-d '{"text":"Hello world","models":["gpt-6.1-sol"]}'
/api/v1/modelsStatus
betaThe API is free and open while in beta. There are no accounts or keys yet, and no published request limits — there is basic abuse protection, but we won't quote a number we can't guarantee. Build against it, but don't depend on it for anything critical until it's out of beta.
Paid tiers with committed limits and an SLA are planned. If you want to be told when they land, or you need higher volume sooner, get in touch.
Response fields
| Field | Meaning |
|---|---|
| tokens | Token count for that model |
| method | exact — counted with the real tokenizer · estimate — character ratio, no public tokenizer exists |
| input_cost_usd | Cost to send this text as input |
| fits_context | Whether it fits that model's context window |
Case study: team cost dashboard
A concept for what per-team LLM cost tracking could look like. Built as a design exercise, not a live product.
Every number in this dashboard is randomly generated sample data. Nothing is connected to a backend, no accounts exist, and no real usage is being tracked. It exists to show the shape of the idea.
The problem it addresses
A calculator answers "what will this cost?" before you send a prompt. It doesn't answer "what did we spend last month, which team spent it, and why is it climbing?" — which is the question that actually reaches a finance or engineering lead once a company has several LLM features in production.
Two audiences, one surface
| View | Answers |
|---|---|
| Business | Spend over time, budget vs. cap, cost by team and project, which provider dominates |
| IT / developer | Request volume, error rate, latency, per-key rate-limit headroom, request log |
What making it real would require
Authentication so each organisation sees only its own data, a database logging every API call, an SDK that wraps a customer's LLM calls to feed that log, and a scheduled job for budget alerts. That's a different product from a calculator — which is why it's shown here as a concept rather than shipped as a feature.
Contact
Token Bench is built by Vector & Value AI. Email is the best way to reach us.
A few details help us answer quickly:
| A price looks wrong | The model name, a link to the provider’s pricing page, the price and currency it shows, and the date you checked it. |
|---|---|
| API access | Roughly how many requests a month you expect. |
| Research or data | Which study or dataset you’re asking about. |
| Something’s broken | The page address, your browser, and what you expected to happen. |
| Business enquiries | A line about your company and what you’re looking for. |
Corrections backed by an official provider page are applied directly, and the “last verified” date in our FAQ is updated to match. Figures on this site are estimates for planning. They aren’t financial or procurement advice, and they don’t reflect negotiated enterprise rates.
FAQ
POST /api/v1/tokenize endpoint returns counts and cost for any text across all supported models. It’s free and in beta: there are no accounts or keys yet, and no published request limits. See the API page for details.