AI token cost calculator for Claude, GPT and Gemini
Type how much you send, how much comes back and how often. See the monthly estimate on eleven current models from Anthropic, OpenAI and Google, with the token counts behind it.
This calculator estimates what an AI job costs a month through the API. Type how many words you send in each request, how many come back and how many requests you make, and it prices the job on eleven current models from Anthropic, OpenAI and Google.
The prices come from each maker's own pricing page, which I last read on 2026-10-28. The sums run in your browser, and nothing you type leaves your device. It starts with an example: answering 3,000 customer emails a month.
Your instructions, the question and any document or chat history sent with it.
The share that's the same every time and cached, such as a long system prompt.
The answer, plus any thinking the model does first. Thinking is billed as output.
About $0.64 a month on GPT-6 Luna, up to $95.07 on Claude Fable 5.1, for 3,000 requests.
| Model | A month | Tokens per request |
|---|---|---|
| GPT-6 LunaOpenAI | $0.64$0.00021 each | 800 in, 267 out |
| Claude Haiku 5.5Anthropic | $0.95$0.00032 each | 1,189 in, 396 out |
| Gemini 3.5 Flash-LiteGoogle | $2.72$0.00091 each | 800 in, 267 out |
| Gemini 3.8 FlashGoogle | $4.80$0.0016 each | 800 in, 267 outFrom January 1, 2027: $9.61 a month at the announced price. |
| GPT-6.1 SolOpenAI | $12.81$0.0043 each | 800 in, 267 out |
| Gemini 4 ArgonGoogle | $12.81$0.0043 each | 800 in, 267 outFrom December 1, 2026: $25.62 a month at the announced price. |
| Gemini 3.1 Pro PreviewGoogle | $14.41$0.0048 each | 800 in, 267 out |
| Claude Sonnet 5.5Anthropic | $19.01$0.0063 each | 1,189 in, 396 out |
| Claude Opus 5.5Anthropic | $38.03$0.013 each | 1,189 in, 396 out |
| GPT-6 AstraOpenAI | $64.05$0.021 each | 800 in, 267 out |
| Claude Fable 5.1Anthropic | $95.07$0.032 each | 1,189 in, 396 out |
Standard API prices from each maker's own pricing page, checked October 28, 2026. An estimate: your bill comes from the token counts the API reports.
How to estimate your AI API cost
- Pick the example closest to your job, or type your own numbers over it.
- In "Sent per request", count everything that goes to the model each time: your instructions, the question, and any document or earlier messages sent along with it.
- In "Written per request", count the answer, plus the thinking. Reasoning models think before they answer, and every maker bills that thinking as output. OpenAI says reasoning tokens "are billed as output tokens", and Anthropic reports thinking as part of "the billed output tokens".
- If part of every request is identical, such as a long system prompt, and you use prompt caching, put that share in "from the cache".
- Tick "Send as batches" if the work can wait. A batch is processed in the background instead of answered straight away, and on every model here its input and output prices are half the standard ones.
The table lists the models from cheapest to most expensive for your job. Each row shows the month, one request, and the tokens the model would count.
How the calculator turns words into tokens
Models are billed per token, a chunk of text that's often part of a word. Each maker publishes a rule of thumb for English, and the calculator uses that maker's rule for its own models.
| Maker | Its rule of thumb | 1,000 words is about |
|---|---|---|
| OpenAI | 1 token is about 4 characters, or 0.75 words | 1,333 tokens |
| 1 token is about 4 characters; 100 tokens is 60 to 80 words | 1,333 tokens | |
| Anthropic | about 3.5 characters on older models; the newer tokenizer makes about 30% more tokens | 1,981 tokens |
Anthropic's rule is in characters, so I turned words into characters with OpenAI's rule (4 characters is 0.75 words, so a word is about 5.3 characters with its space). That makes the Claude figure an estimate built on two rules of thumb. Anthropic says Claude 4.7 and later models use the newer tokenizer, which covers all four Claude models here.
The difference matters when you compare. The same 1,000 words count as more tokens on Claude than on GPT or Gemini, so the price per million tokens doesn't compare models on its own.
If you know your token counts, switch "Count in" to Tokens. The calculator then uses your numbers as they are for every model. A count measured on one maker's model will be off for the others.
A worked example
The first example is 600 words in, 200 words out, 3,000 times a month. On Claude Opus 5.5, that's about 1,189 tokens in and 396 out per request.
Input costs $4 per million tokens: 1,189 tokens is $0.0048. Output costs $20 per million: 396 tokens is $0.0079. One email costs about $0.0127, and 3,000 of them about $38 a month.
The answer is a third as long as the request, and it still costs more. Output is priced five times higher than input on every model here except Gemini 3.1 Pro Preview (six times) and Gemini 3.5 Flash-Lite (more than eight times). For most jobs, the length of the answer changes the bill more than the length of the prompt.
The prices it uses
Standard prices in US dollars per million tokens, for prompts up to each model's long-prompt threshold, as published on 2026-10-28.
| Model | Input | Output | Cache read |
|---|---|---|---|
| Claude Fable 5.1 | $10 | $50 | $0.25 |
| Claude Opus 5.5 | $4 | $20 | $0.20 |
| Claude Sonnet 5.5 | $2 | $10 | $0.10 |
| Claude Haiku 5.5 | $0.10 | $0.50 | $0.01 |
| GPT-6 Astra | $10 | $50 | $1 |
| GPT-6.1 Sol | $2 | $10 | $0.10 |
| GPT-6 Luna | $0.10 | $0.50 | $0.01 |
| Gemini 4 Argon | $2 | $10 | $0.10 |
| Gemini 3.1 Pro Preview | $2 | $12 | $0.20 |
| Gemini 3.8 Flash | $0.75 | $3.75 | $0.075 |
| Gemini 3.5 Flash-Lite | $0.30 | $2.50 | $0.03 |
Three things change these prices, and the calculator applies all three:
- Long prompts. Some models charge a higher rate once the prompt passes a size, and the calculator applies it to the whole request: Claude Haiku 5.5 over 100,000 tokens, the three GPT-6 models over 272K and Gemini 3.1 Pro Preview over 200K. The other Claude models charge the same rate up to their full 1M-token context window.
- Dated changes. Google has announced that Gemini 4 Argon doubles in price on December 1, 2026, and Gemini 3.8 Flash on January 1, 2027. The calculator uses the price for today's date and shows the new monthly figure under those rows.
- Batches. Every model here has batch input and output prices at half the standard ones. On Claude, the cache discount and the batch discount stack. On Gemini 3.1 Pro Preview, a cache read costs the same in a batch.
What the estimate leaves out
The calculator doesn't count cache writes, which cost more than ordinary input: 1.25 times on Claude for a 5-minute cache, and 1.25 times on OpenAI's GPT-5.6 and later for the default 30 minutes. Google also charges for cache storage by the hour on most Gemini models. With caching on, your bill will be higher than the estimate.
It doesn't count tools either, such as web search, which are billed on top of tokens. On the Claude API, web search is $10 per 1,000 searches. On Gemini, grounding with Google Search is free for 5,000 requests a month, shared across Gemini 3 and newer models, then $14 per 1,000.
It leaves out regional and fast processing. Keeping processing in one region costs more: 1.1 times on Claude with US-only inference, and 10% more on OpenAI's regional endpoints for models released since March 5, 2026. Fast modes cost more too; Claude Opus 5.5 in fast mode is $8 in and $40 out per million tokens.
Images, audio and video are priced differently, and the calculator covers text only.
A monthly plan for a chat app isn't billed per token, so this calculator doesn't apply to it.
For an exact count before you send, the APIs have token-counting endpoints. Anthropic's is free to use, subject to rate limits. After you send, each response reports the tokens it used.
How I keep the prices current
The prices live in one data file with the date I last checked them, and the calculator prints that date under the table. My weekly audit re-reads the three pricing pages and updates the file when a price has changed. I describe that routine in how this site runs itself.
I chose one or more current models from each maker's main family and left out older models, image, audio and video models, and models on a limited-time promotional price. If a price here doesn't match the maker's page, the maker's page is right, and I'd like to hear about it.
Your numbers stay on your device
The calculator is a small script that loads with this page. It makes no network requests after that, and it stores nothing, so a reload starts from the example again.
It needs JavaScript to work out your own numbers. Without it, the page shows the first example's table and this explanation.
I checked the sums against 34 cases I worked out by hand, including the long-prompt thresholds (a prompt exactly at the threshold pays the lower price), batch and cache prices, the dated changes on the day before and the day of each change, and how amounts are rounded.