Inference usage

Inference usage pays for model calls provided by Warp. Its cost depends on the model’s rates and the work required to complete your request. See Usage and billing for inference rates and how they differ from compute and platform charges.

How inference usage is calculated

• Model choice - Models have different input, output, and cache prices. A lower-cost model can reduce spending without reducing the number of tokens.
• Input and output - Your prompt, conversation history, attached context, and generated response contribute to token usage.
• Caching - Cached input can cost less than new input. Cache-read and cache-write tokens are reported separately.
• Task complexity - A task can require several model calls, including calls made while using tools or summarizing a long conversation.
• Provider tools - Features such as web search can add charges beyond token costs.

Two similar prompts can use different amounts. Use the reported breakdown to compare tasks rather than treating a prompt as a fixed-price unit.

For example, two requests with the same token count can cost different amounts if they use models with different prices or one request uses cached input.

Customer-supplied inference

If you use your own API key or inference endpoint, your provider bills that inference separately. Bills from your own inference provider are not fully represented in Warp’s usage totals. Separate platform charges can still apply.

Related pages

• Usage and billing - Inference, platform, and compute rates
• Agent model choice - Supported models and model selection
• Use tokens more efficiently with AI coding agents - Reduce inference usage