Configuring the pricing catalog

AI Observatory never downloads provider pricing. Define estimates in the published configuration:

'pricing' => [
    '_meta' => [
        'version' => 'internal-2026-07',
        'effective_date' => '2026-07-01',
    ],
    'provider-name' => [
        'model-name' => [
            'input_per_million' => 10,
            'output_per_million' => 30,
            'cached_input_per_million' => 1,
            'cache_write_input_per_million' => 12.5,
            'currency' => 'USD',
        ],
    ],
],

Provider and model keys must match the normalized values stored on the span. Rates are amounts per one million tokens.

Input usage is required. Output usage and its rate are required for operations where output tokens apply; input-only operations such as embeddings can still be estimated. Cached-read and cache-write rates are required only when their corresponding usage is greater than zero.

Provider adapters normalize uncached input, cache reads, and cache writes into independent categories before calculation. The currency must be a three-character code. Negative or contradictory usage, invalid rates, or incomplete entries produce no estimate.

The catalog version and effective date are stored in span pricing metadata. Every displayed amount is an estimate.