Pricing

How tokens are counted

Learn how input and output tokens are counted and billed, how the cached-prompt discount works and how to estimate a request's cost before you send it.

Tokens are how Claude (and every other LLM) meters work. Roughly 4 characters of English text — and roughly what you pay for.

All models, capabilities, and prices Documentation index for AI agents: https://oneprovider.dev/docs/llms.txt

Rule of thumb

1 token ≈ 4 English characters ≈ 0.75 words.
A page of prose (~500 words) is ~670 tokens.
Source code, non-Latin scripts, and emoji tokenize denser — expect 1.2–2× the English rate.
A 200k-token context window is roughly 150k English words — around 300 pages.

Four kinds of tokens on your bill

Reference data

Kind	What it is	How it’s billed
Input	Every token the model reads: system prompt, prior turns, current user message, tools schema, images.	Full input rate.
Output	Every token the model writes: response text, tool calls, thinking (if enabled).	Full output rate (typically 5× input).
Cache read	Input tokens served from a prior request’s cache breakpoint.	10% of the input rate for most models; exact rates per model are in the models table .
Cache write	The first request at a cache breakpoint — establishes the cache for later reads.	125% of the input rate for most models; exact rates per model are in the models table .

Every response carries these counts in the usage field — see Usage & billing .

Estimating before you send

The Anthropic SDKs expose a count-tokens helper you can run locally:

python

from anthropic import Anthropic
client = Anthropic(base_url="https://api.oneprovider.dev")
count = client.messages.count_tokens(
model="claude-sonnet-5-5",
messages=[{"role": "user", "content": "Your long prompt here"}],
)
print(count.input_tokens)

The count is a lower bound — it doesn’t include the output the model is about to write. For budgeting, add a max_tokens ceiling and assume the worst case.

How to estimate cache savings

There is no universal multiplier across models. For the current balanced model claude-sonnet-5-5 or any other model, use its separate Cache read and Cache write values from the generated rates table . Compare them with the selected model's input rate and expected number of repeated reads.

See the live calculator /pricing has a slider that computes monthly cost across model + cache-ratio + input/output mix — same numbers as the Rates table, just easier to play with.

Quick answers

Which base URL do I use?

Claude Code and Anthropic SDKs use https://api.oneprovider.dev. OpenAI SDKs and OpenAI-compatible clients use https://api.oneprovider.dev/v1. One key works behind both; the model list is on the models page ; rates are on the pricing page .

How do I authenticate?

Send the raw sk-… key in x-api-key or Authorization: Bearer: one of them, never both. Keep the key in an environment variable, not in a file that gets committed. More on authentication .

Where do I get a key?

Open your cabinet in the Telegram bot @oneprovider_robot or at app.oneprovider.dev, top up the balance and press “Issue a key”. Each key can carry its own spend limit. Cabinet & API keys .

What do 401, 404 and 429 mean?

401: the key is missing, mistyped or sent in the wrong header; paste the whole key without quotes or spaces. 404: usually a wrong base URL, for example a missing or doubled /v1. 429 with “Insufficient balance”: the cabinet balance or the key’s spend limit is used up, so top up instead of retrying; any other 429 is a rate limit, so wait for Retry-After. Details: error codes and troubleshooting .

Why do streams cut off or reset?

VPN and VLESS proxies often buffer SSE or break HTTP/2 negotiation. OneProvider has no regional restrictions and does not buffer streams, so turn the proxy off and retry.

Where do I ask if this is not enough?

Message @Oneproviderhelpbot on Telegram or open “Support” in the web cabinet, with the HTTP status code and the exact error message. How support works .

Last updated: 2026-10-05 Found a bug? @Oneproviderhelpbot