Model Cost Map

LiteLLM prices every request from its model cost map, model_prices_and_context_window.json. Each entry maps a model to its rates, context limits, and capabilities. The proxy fetches the latest map from GitHub at startup, so new models and price changes arrive without upgrading LiteLLM. To change a price for your own deployment, see Custom LLM Pricing and Custom Model Cost Map.

Every price is USD per billable unit

Every price in the map is in US dollars for one unit of what the provider bills. For token-priced models that unit is a single token, so "input_cost_per_token": 1e-05 means $0.00001 per input token, which is $10 per 1M input tokens. 1e-05 is just how Python prints 0.00001.

Key	Unit	Example
input_cost_per_token, output_cost_per_token	USD per token	3e-06 is $3 per 1M tokens
cache_read_input_token_cost, cache_creation_input_token_cost	USD per cached token read or written	3e-07 is $0.30 per 1M tokens
output_cost_per_image	USD per generated image	0.04 is $0.04 per image
input_cost_per_second, output_cost_per_second	USD per second of audio or video	0.1 is $0.10 per second
input_cost_per_query	USD per search or tool query	0.005 is $5 per 1,000 queries
input_cost_per_character	USD per input character	1.5e-05 is $15 per 1M characters

Variants of a rate keep the same unit and add a suffix. _batches, _priority and _flex are the batch and service-tier rates, _above_200k_tokens and _above_272k_tokens are long-context rates that apply once the prompt crosses that size, and cache_creation_input_token_cost_above_1hr is the 1-hour prompt cache write rate.

To convert a vendor's "$X per 1M tokens" price into a cost map value, divide by 1,000,000. For example, $2.50 per 1M tokens is 2.5e-06.

model_list:
  - model_name: my-model
    litellm_params:
      model: openai/gpt-6.1-sol
    model_info:
      input_cost_per_token: 2.5e-06   # $2.50 per 1M input tokens
      output_cost_per_token: 1.5e-05  # $15 per 1M output tokens

How a request is priced

After the response comes back, LiteLLM reads the provider's usage block, splits it into the token types that bill differently, and multiplies each count by its rate. Take an entry with input_cost_per_token: 1e-05, cache_read_input_token_cost: 2.5e-07 and output_cost_per_token: 5e-05. A request with 112,000 prompt tokens, 100,000 of them cache reads, and 800 output tokens costs:

 12,000 uncached input tokens x 1e-05    = $0.12
100,000 cache read tokens     x 2.5e-07  = $0.025
    800 output tokens         x 5e-05    = $0.04
                                  total  = $0.185

The same pattern covers everything else. A prompt over a long-context threshold uses the _above_..._tokens rate, a request sent with service_tier: priority or flex uses the matching suffixed rate, and audio, image, video, and reasoning tokens use their own rate when the entry has one and the text rate when it does not. Entries with a regional uplift, such as a data residency region, multiply the total by the regional_processing_uplift_multiplier_<region> stored on the entry.

Some providers return the cost of the request themselves, and LiteLLM uses that number. xAI reports cost_in_usd_ticks, where one tick is 1e-10 USD. Since xAI reports one total, LiteLLM splits it between input and output in the ratio its own token math gives.

Currency

Spend is always tracked in USD. LiteLLM does no currency conversion at request time. When a vendor publishes prices only in another currency, the cost map entry is converted to USD once at a fixed rate, and that rate is recorded in the entry's metadata.comment.

The result of each request is written to LiteLLM_SpendLogs.spend as a USD float and rolled up into key, user, and team spend for budgets. Because it is a float, you may see values such as 0.14500000000000002 in logs. The tail is far below a cent, even summed over billions of requests.

When a request costs $0

A request prices to $0 when its model, or a rate the request used, is missing from the map. The request is still logged with spend = 0. When the model has a non-zero rate on its entry but the request still prices to $0, the proxy also counts it in the litellm_zero_cost_requests_total Prometheus metric. See Requests that price to $0 for how to read the warning and fix the missing rate.