AI token calculator: tokens and API cost per model

Every LLM token calculator has to answer one question first: whose tokenizer? Each AI company splits text its own way, so the same paragraph is a different number of tokens on each model. This one uses the best method each company allows and labels every number with it.

• OpenAI models: exact count. As an OpenAI token calculator it runs OpenAI's o200k_base tokenizer in your browser, through the open-source gpt-tokenizer library (version 4.0.0). OpenAI's own tokenizer library, tiktoken, maps GPT-4o, GPT-4.1 and every GPT-5 model to o200k_base.
• GPT-6 models: the same tokenizer, not yet confirmed. tiktoken has no GPT-6 entry on October 28, 2026, and OpenAI has not said which tokenizer GPT-6 uses. A public comparison of 31 texts, run on September 26, 2026, found GPT-6 Astra, Sol and Luna reporting exactly the same input counts as GPT-5.5, so we count them with o200k_base and label them that way.
• Claude models: our measured estimate. Anthropic offers no tokenizer that runs offline, so the Claude token calculator divides your characters by the ratio we measured for that kind of text (table below). Anthropic's pricing page says Claude 4.7 and later models share one newer tokenizer, and in our test Haiku 5.5, Sonnet 5.5 and Opus 5.5 agreed within 4 tokens.
• Gemini models: Google's rule. As a Gemini token calculator it uses Google's own rule of thumb of about 4 characters per token. We did not measure Gemini, because that needs a Google API key.
• Other models (Grok, DeepSeek, Qwen, Mistral, Kimi and the rest): there is no tokenizer for them in the browser, so the calculator shows the OpenAI count as a stand-in and labels it.

How we measured Claude

On October 28, 2026 we sent six texts through Claude Code 2.1.303 with Claude Haiku 5.5, from an empty folder with every tool switched off, and read the input tokens in its JSON output. The same request without the text (a fixed one-line instruction) came to about 4,008 tokens, measured four times. Subtracting it leaves the tokens of the text itself. The English text and the code ran twice and matched within 5 tokens, and Claude Sonnet 5.5 and Opus 5.5 got the same two texts as a cross-check.

Claude needed 1.28 to 1.78 times as many tokens as OpenAI for the same text: the fewest extra for the table, the most for Portuguese. That fits the public comparison above, where Claude Opus 5.5 counted 53.9% more tokens than GPT-5.5 across its 31 texts. One sample per kind of text gives a rule of thumb, not a guarantee: for an exact Claude count, Anthropic's token counting endpoint is free to use with an API key.

Nothing you paste leaves the page. The counting and the cost math run in your browser, the tool is masked from our page analytics, and the only file it downloads is the OpenAI tokenizer (about 1 MB), from our own site.

The capture below is this calculator with our English test text (1,316 characters and 239 words; the table above counts 2 more characters for the line break the test added), the six models it opens with, a 300-word reply and 100 requests a day. OpenAI's tokenizer counted 268 tokens, our Claude estimate came to 409 and Google's rule to 329 for Gemini. GPT-6 Luna was cheapest at $0.000195 a request ($0.584 a month); Claude Opus 5.5 cost the most at $0.0119 a request ($35.81 a month), about 61 times as much.

These are list prices in US dollars per 1 million tokens, input and output, for the 26 language models in the calculator, each from the vendor's own pricing page (one exception: Meta prices no API for Muse Spark Open, so its row shows the endpoint price Artificial Analysis lists). Most were read for our AI model leaderboard on October 18, 2026; on October 28, 2026 we re-read the OpenAI, Anthropic and Google pages and they still matched. Output costs 2 to 6 times as much as input on every model here, so a long answer costs more than a long prompt.

The calculator uses list prices only. Batch processing is half price at Anthropic and at OpenAI, whose pricing page lists batch rates at half the standard ones for the GPT-6 models, and cached input costs a fraction of the list price, so a real bill can be lower.

Five habits do most of the work, and each one shows in the calculator: change the reply length, the model or the text and watch the monthly figure move.

• Cache what every request repeats. If your requests start with the same instructions or document, prompt caching bills the repeat at a fraction of the input price: 5% on Claude Opus 5.5 and Sonnet 5.5 (Anthropic), and $0.10 instead of $2 per million tokens on GPT-6.1 Sol (OpenAI). Claude Code already does it: even with its tools switched off, it sent 3,200 to 4,000 tokens of its own instructions with each one-line request in our test, and after the first request most of them were counted as cache reads.
• Batch what can wait. Reports, tagging and summaries that can arrive within hours cost half as much through Anthropic's Batch API and OpenAI's batch rates.
• Test the smaller model first. Claude Haiku 5.5 lists $0.10 / $0.50 per million tokens against $4 / $20 for Claude Opus 5.5, 40 times less. Run a handful of your real requests on both before you settle on the big one.
• Stay under the long-prompt line. Claude Haiku 5.5 charges 5 times as much for prompts over 100,000 tokens, and the GPT-6 models charge double for input and 1.5 times for output above 272,000 input tokens. Paste the part of a document the task needs, not the whole file, and trim old chat history.
• Ask for shorter answers. Output costs 2 to 6 times as much as input, so say the length you want ("three bullet points") and set a maximum output in the API. Models that think before answering bill that thinking as output too: Google's pricing page labels its output price "including thinking tokens".

If you use Claude Code on a subscription rather than the API, tokens show up as usage limits instead of a bill; our guide to Claude Code usage limits explains the 5-hour and weekly caps and how to make them last.

• AI model leaderboard: which model leads at chat, coding, writing and images, with the same prices and a source for every score.
• Claude Code pricing: Pro, Max, Team and API prices side by side, and which plan to pick.
• Claude Code usage limits: how the 5-hour and weekly caps work and ways to save tokens.
• Best AI coding agents: 12 coding agents compared with benchmark scores and prices.

We teach Claude Code in our own course, linked below. This page has no affiliate links, and every price links to the vendor's own page.