Phoenix can update model prices between releases

Arize’s announcement describes default behavior while automatic downloads are enabled and external resources allowed: Phoenix checks for newer prices shortly after startup, then roughly daily, in the background. Updated rates apply to newly received model-call records, called spans, after the pricing cache refreshes. Existing trace costs stay unchanged, and custom prices retain priority over matching built-in rates.

The cost guide identifies two settings that disable downloads: PHOENIX_DISABLE_MODEL_COST_AUTO_UPDATE=true or PHOENIX_ALLOW_EXTERNAL_RESOURCES=false. These deployments receive built-in catalog changes through Phoenix upgrades. Calculations require input, output and total token counts, plus model and provider information. Supported OpenInference instrumentation captures these automatically; manually instrumented code must supply the attributes. Phoenix combines usage with model prices. In Settings > Models, teams can set separate input and output rates, including for models absent from the built-in table.

For readers starting locally, the installation guide gives a concrete sequence: install the arize-phoenix package with pip, then run phoenix serve. Open the browser link printed in the terminal; the default interface uses port 6006 on the local machine. That starts the dashboard for exploring logs and traces. Sending traces from an application is the next step after this local setup.