LLM Token-to-Cost Calculator (2026) — GPT-4o, Claude, Gemini & 13 More Models

Token pricing is the closest thing the AI industry has to a unit-economics number, and yet most developers run their app for weeks before they look at it. Paste a prompt below and see what one call — or a million of them — would cost across 16 models.

How AI token pricing actually works

Every major LLM provider charges per token — a fragment of a word that the model’s tokeniser produces when it reads or writes text. A rough rule of thumb that holds across OpenAI, Anthropic, and Google tokenisers is roughly four characters per token, or about three-quarters of a token per word for English prose. Code, JSON, and non-Latin scripts tokenise more densely (more tokens per character), so a calculator like this gives you a working estimate, not a final invoice.

Almost all vendors price input tokens (what you send in: system prompt + user message + chat history + retrieved context) and output tokens (what the model writes back) at different rates. Output is typically 3–5× more expensive than input. For chat-style usage the cost is dominated by input. For long-form generation, it flips.

How to use the calculator

1. Paste a representative prompt into the text box, or switch to the “Enter counts” tab if you already know your token counts.
2. Set the expected output tokens — how long the model’s response usually is. A chat reply is ~200, a summary ~400, a long explanation ~1,000+.
3. Scale it up using the calls-per-period control. “100 calls per day” projects a monthly bill; “1 call” gives you a single API-call price.
4. Pick a quick preset if you want a starting point: chatbot reply, document summary, code generation, or a typical RAG question.
5. Sort by total cost to see the cheapest model first, then re-sort by vendor or input price to compare a specific family of models.

Realistic cost examples

Chatbot — short reply, light context

~600 input tokens (system prompt + last few turns) · 200 output tokens. Per-call cost on Claude Haiku 3.5 is about $0.0013; on GPT-4o it’s $0.0035; on Claude Opus 4 it’s $0.024. Sounds tiny — multiply by 100k calls and the gap becomes $130 vs $2,400 a month for the same workload.

Document summary — long input, short output

~4,000 input tokens (a long article) · 400 output tokens (a tight summary). This is where the input/output asymmetry bites: input dominates. Gemini 1.5 Flash runs about $0.0004/call here, GPT-4o about $0.014, Claude Opus 4 about $0.090. Models with cheap input win hard on summarisation.

Code generation — medium input, long output

~1,500 input tokens (instructions + file contents) · 1,000 output tokens (refactored file). Output dominates. Claude Sonnet 4 lands around $0.020/call; GPT-4o around $0.014; the smaller models trade quality for an order of magnitude saving.

RAG question — huge input, small output

~8,000 input tokens (retrieved chunks + question) · 250 output tokens (concise answer). The classic enterprise pattern. Input cost is everything; pick a model with cheap input or trim your retrieved context, or both. Gemini 1.5 Pro’s long-context pricing tier matters here.

How to reduce your token bill

• Use the cheapest model that meets the bar. Many tasks don’t need a flagship. Run an eval comparing Haiku 3.5 / GPT-4o mini / Gemini Flash against your current model on a hundred real examples — you’ll often find the cheap model is fine.
• Trim the system prompt. A 2,000-token system prompt sent on every call is a $20 bill per million calls on GPT-4o that you may not even need. Move static instructions to the model’s training, fine-tuning, or a tighter, shorter prompt.
• Cap output length. Set max_tokens deliberately. Letting a verbose model ramble for 1,000 tokens when 200 would do is a 5× cost multiplier on the output side.
• Use prompt caching. Anthropic, OpenAI, and Google all now offer cached-input pricing at 50–90% off for repeated prompts — huge for chat apps with long stable system prompts.
• Batch where latency doesn’t matter. Most providers offer 50% off via a batch API for jobs you can wait minutes/hours for (overnight enrichments, evals, classification).
• Route smartly. Send easy queries to a cheap model and escalate only the hard ones to a flagship. Even a simple keyword-based router can cut blended cost 60–80%.

Frequently asked questions

The bottom line

If you’re building anything serious on top of LLM APIs, the token-cost spread between models is the single biggest variable in your unit economics. Before you ship, run the cheapest plausible model through a real eval and compare it to the flagship you’re defaulting to. The savings compound from the day you switch — and the calculator above is meant to be the back-of-the-envelope you reach for every time someone in a meeting asks “what would that cost in production?”