LLM model catalogue: prices, context windows and capabilities

How is this catalogue kept current?​

DVARA reviews every price weekly before it is published. Each check marks a model for a closer look: a cached-input price outside 10–50% of the input price, output cheaper than input, a zero price on a hosted model, a missing cache price where the provider bills caching, a context window smaller than the output limit, and a price more than 1% away from another published price list. A mark never hides a model. Each model page shows its marks.

How does DVARA use these prices?​

In DVARA, if you haven't set your own price for a model, its calls are costed at the price in the catalogue that ships with your DVARA version. A price you set in Flightdeck always wins. You can switch catalogue prices off. Then a model with no price you set is counted as costing nothing. Every cost record says which price it used.

Each DVARA release packages a snapshot of this catalogue, so an install with no internet access has it. Flightdeck can also read your own copy of models.json from a URL you choose. It refuses a copy that is damaged, much smaller than the packaged one, or fails a price check it does not flag, and keeps the last good copy. Read more about LLM cost attribution and budgets.

What do the columns mean?​

• Input and output are the list prices per million tokens sent to and returned by the model.
• Cached input is the price per million input tokens read from the provider's prompt cache.
• Context is the most tokens the model accepts in one call, input and output together.
• Max output is the most tokens the model returns in one call.
• Vision means the model accepts images. Structured output means it can be held to a JSON schema.