Metrum AI Router

Metrum website: metrum.ai. See the Metrum AI Router product page for product information and enterprise inquiries.

For v4.3.0 operations, see the upgrade guide, AWS Bedrock Converse upstream, usage provenance, and unbilled count-token requests. The fine-tuning recipe links measured evidence and promotion limits; it is not a default model change.

Agent workloads often make dozens of model calls per task. Paying frontier price for each call is the default path, and token spend shows up after the monthly budget is gone. In one committed Harbor run, Codex on big-coder used 48,470 Harbor input tokens in a single deterministic job…