Skip to main content
The Guild platform computes estimated LLM spend in real-time based on industry-standard per-million-token list prices. All calculations are handled linearly and accurately at the individual LLM call level.

​
Effective-dated pricing

LLM cost calculations are date-aware. Guild prices each usage day at the rate in force on that specific day, so updating a model’s list price does not retroactively alter historical usage dashboard figures. A model can define a single flat rate or a list of price periods, each with an effective_from date (at 00:00 UTC). Spend lookups select the price in force on the day the tokens were consumed. For example, the claude-sonnet-5 promotional rate applies through August 31, 2026, and a higher rate applies from September 1, 2026 onward. Some OpenAI models, such as gpt-5.6-terra and gpt-5.6-luna, also use effective-dated rates: their launch rates apply through August 6, 2026, and reduced rates apply from August 7, 2026 onward.

​
Serving platform rates

The same model can reach Guild through more than one platform, such as the publisher’s own API, Bedrock, or OpenRouter, and a platform can charge a different rate than the publisher. When the pricing table has a rate for the platform that served a call, Guild prices the call at that platform’s rate. Otherwise it uses the publisher’s list price. A platform rate can also be specific to a region, such as Bedrock in eu-west-2; a call served from a region without its own rate uses the platform’s general rate. In the Models breakdown, a model that a gateway served shows a via [Platform] marker, and its rate panel names the platform whose rates it shows.

​
Default list prices

Below are the default list prices (in USD per Million Tokens) configured on the platform.
The claude-sonnet-5 launch rate is promotional. It applies through August 31, 2026; from September 1, 2026 the rate rises to 3.00inputand3.00 input and 15.00 output per million tokens.

​
Default fallback rate

If a model name is not recognized or does not match any of the custom tiers above, the system logs a warning and falls back to the default list rate (matching the Claude Sonnet 4.6 tier). The Models breakdown table displays a warning indicator next to the model name to signal that the spend is an estimate.
  • Input: $3.00 per Million
  • Output: $15.00 per Million
  • Cache Read: $0.30 per Million
  • Cache Write: $3.75 per Million

​
Prompt caching dynamics

To provide highly accurate cost accounting, prompt caching is calculated uniquely per provider:
  • Anthropic / OpenAI: Prompt cache write tokens are priced at a premium rate (cache_write / cache_create), and subsequent hits are charged at a heavily discounted cache_read rate.
  • Google Gemini: Google structures prompt caching differently, billing cache storage per hour rather than a per-token write rate. As a result, cache_write_tokens are priced at $0.00 (cache_create: 0.0), while cache_read_tokens represent the discounted input rate.
  • Double-charging prevention: The input_tokens count (labeled Prompt tokens in the dashboard) is canonically cache-read inclusive for all providers, so it already contains the cache_read_tokens. To prevent double-billing, the platform’s query layers deduct cache_read_tokens from input_tokens on each LLM call before applying the pricing rates, isolating the billable uncached prompt portion. Thus, the calculation always evaluates as billable_input = max(input_tokens - cache_read_tokens, 0).