ChatGPT and OpenAI token counter
OpenAI publishes the encodings for GPT-4o, GPT-4.1, and o3, so this counter runs the real tokenizer in your browser and those counts are exact as you type. GPT-5.5's encoding is not published yet, so its count is an estimate, marked as such.
0 words · 0 characters
GPT models at a glance
Your text counted against every current GPT model. Counting against every provider at once? Use the full token counter, or see every model's limit in the context-window comparison.
| Model | Your text | Context window | Max output | Input $/M tokens |
|---|---|---|---|---|
| GPT-5.5 very long prompts bill at a higher tier | — | 1.1M ~788K words | 128K | $5 |
| GPT-4.1 | — | 1.0M ~786K words | 33K | $2 |
| GPT-4o | — | 128K ~96K words | 16K | $2.5 |
| o3 | — | 200K ~150K words | 100K | $2 |
Windows and prices verified against provider documentation on 2026-07-07.
GPT token questions, answered
What is ChatGPT's token limit?
It depends on the model behind it: GPT-4o reads 128,000 tokens at once, while GPT-5.5 and GPT-4.1 take roughly a million. The ChatGPT app also enforces its own smaller, unpublished in-app windows, so long conversations quietly fall out of scope earlier than the model's real limit.
Are these counts exact?
For GPT-4o, GPT-4.1, and o3, yes: OpenAI publishes their encodings and the counter runs the real tokenizer locally. GPT-5.5's encoding is not published, so its count is an estimate, marked with a tilde.
How many tokens is a word for GPT models?
English prose averages about 1.3 tokens per word on the o200k encoding, so 100 tokens is roughly 75 words. Code and other languages tokenize less predictably, which is why a real count beats the rule of thumb.
Do long ChatGPT conversations use more tokens?
Yes. The app resends the visible conversation with every message, so message thirty carries the previous twenty-nine. That is why limits feel tighter late in a thread and why controlling the context beats resending all of it.
Counting tokens is managing context by hand. NodePad is a canvas for doing it on purpose: fork side questions, keep every path, and choose exactly what GPT (or any other model) sees.
Try NodePad free