Updated July 2026. What a token is, what it costs per million across GPT, Claude, Gemini and DeepSeek, why output is pricier than input, and how to estimate your own bill in under a minute.
Every large language model API charges the same way: per token. Not per request, not per word — per token. If you have ever looked at a pricing page and seen "$3.00 / 1M tokens" and wondered what that means for your actual bill, this guide breaks it down with real 2026 numbers. When you want an exact figure for your own prompt, paste it into the calculator and it prices input and output separately for any model.
A token is a chunk of text the model reads and writes in. It is not a word and not a character — it sits in between. A rough rule for English: 1 token ≈ 4 characters ≈ 0.75 words. So 1,000 tokens is about 750 words, and this paragraph is roughly 60 tokens. Punctuation, spaces and word-pieces all count. Non-English text is far less efficient: Cyrillic, Chinese, Japanese and Korean routinely use 2–4× more tokens for the same meaning, which quietly multiplies your cost — we cover that in the Cyrillic tax.
Because a single token costs a tiny fraction of a cent, vendors quote prices per million tokens (often written /1M or /MTok). A rate of "$5.00 / 1M input" means one million input tokens cost $5, so a single token costs $0.000005. You will always see two numbers, because input and output are priced separately:
On nearly every frontier model, output is priced about 5× higher than input. Claude Opus 4.8 is $5 input / $25 output. Gemini 3.1 Pro is $2 / $12. The reason is compute: reading your input is a single pass, but generating a reply runs a full forward pass of the model for every token it produces. Output is the expensive direction — which is why capping max_tokens and prompting for concise answers is the easiest way to cut a bill.
Here is where the market sits today, from the cheapest floor to flagship quality. See every current rate on the all-models page.
The spread from floor to flagship is over 60× on output. That is why "which model" is one of the biggest cost decisions you will make — and why comparing two candidates side by side, like on the compare pages, pays off.
Say you send a 2,000-token prompt and get a 500-token answer from Claude Opus 4.8 ($5 / $25 per million):
That looks trivial — until you run it 100,000 times a month, which is $2,250. At scale, the per-token rate is everything. Run the same call through Gemini 2.5 Flash-Lite and the monthly cost drops to about $40. The calculator shows both the per-call and at-scale numbers instantly.
max_tokens and asking for terse answers is free money.Prices are quoted per million tokens. The cheapest models run about $0.10 per million input tokens (roughly $0.0000001 per token); flagship models like Claude Opus 4.8 cost $5 input and $25 output per million. One token is about 4 characters or 0.75 words of English.
Generating output runs a forward pass of the model for every token produced, which is more compute-intensive than reading input. On most frontier models output is priced about 5× higher.
It ranges from about $0.10 (Gemini 2.5 Flash-Lite, GPT-4.1 Nano input) to $25 or more (Claude Opus 4.8 output). Most apps mix input and output, so the effective cost lands in between.
Multiply input tokens by the input rate and output tokens by the output rate, both per million, then add them. The gpt-cost.com calculator does it automatically for any model.
Prices synced from each vendor's official pricing page, July 2026. For exact billing, use each provider's official tokenizer.