Uncensored AI Atlas
Menu

Tool

LLM API token cost calculator

Pick a published rate or enter your own, describe one typical request, and get the cost per request, per day and per month. The formula below the result uses your numbers.

Names the exact Hugging Face repo: coder3101/gemma-4-26B-A4B-it-heretic
Everything you send: system prompt, history, documents, question.
Prompt prefix the provider reads from cache. Counted once.
Include reasoning tokens for thinking models; most providers bill them as output.

Per month

$67.20

$0.00224per request
$2.24per day
$2.24per 1,000 requests

Formula with your numbers

uncached input  4,000 / 1M × $0.40 = $0.0016
cached input    0 / 1M × $0.40 = $0.00
output          800 / 1M × $0.80 = $0.00064
per request     $0.00224
per month       $0.00224 × 1,000 requests × 30 days = $67.20

Rate published by Featherless, checked 14 November 2026: Featherless public model list (/v1/models). Developer plan: billed per token from monthly credits (from $50 a month). The listed context may differ by plan.

Rounded for display; the calculation keeps full precision. USD. Taxes, minimum top-ups and plan fees are not included.

How the calculation works

Providers publish rates in US dollars per 1 million tokens, separately for input (what you send) and output (what the model writes). For one request:

cost = (input − cached) / 1,000,000 × input_rate
     +  cached          / 1,000,000 × cached_rate
     +  output          / 1,000,000 × output_rate

Cached tokens are part of the input, so they are subtracted from the standard input before being priced at the cache rate; nothing is counted twice. If a provider lists no cache rate, cached tokens are priced at the normal input rate and the result says so. If a rate the calculation needs is missing, there is no total. Monthly cost is the per-request cost times requests per day times days.

Getting realistic token counts

  • Read the usage object in real API responses: it reports the exact prompt and completion tokens the provider billed.
  • Thinking models can write thousands of reasoning tokens before the answer. Measure a few real requests before trusting an estimate.
  • Agents make several requests per user action, each resending the conversation. Count requests, not user messages.

Your settings are kept after the # in the address, so you can bookmark or share them. To compare one workload across every provider at once, use the API price comparison; for the rates themselves, see the price table.