A brief for
model providers.

Odin routes its own work across many models and brokers. The Odin R&D lab measures what that costs and how providers behave, and publishes each result with its method and data, whatever it shows. This page is meant to be read in a couple of minutes, printed or forwarded. The longer version is the page for model providers.

What the lab offers.

  • A place in a public, pre-registered benchmark when your models or routing fit the question, under the same decision rules as other providers included in that run. The question, the measure, the decision rule and the spend bound are published before any counted call.
  • A written case study: a field note with the method, the result and its data, on this site.
  • A conversation about Odin's own routing. Any volume would be agreed in writing; none is promised here.

Wording clarification · . This brief previously promised a benchmark under universal provider rules. A published pre-registration sets decision rules for providers included in that particular run; this page does not establish identical rules across all providers or guarantee inclusion. See the publication standard.

What the lab asks.

  • Research credits for the experiments that include you. Credits never buy a result.
  • Research or volume pricing, stated in writing.
  • Consent where your terms ask for it. We name a provider next to a measured number only when its published terms allow it; until then it is shown by letter.

What we have measured.

Each number below is read from the published experiment record it links to, when the site is built.

  1. We sent identical requests for the same model through the brokers we use, at the same moment: 720 counted calls. For claude-sonnet-5-5, the median billed price was 1 times the list price through Broker A, 1.1 through Broker B and 1 through Broker C. For gpt-4o-mini it was 1, 1.1 and 1. No broker answered with a different model than the model asked for: 0 silent substitutions for claude-sonnet-5-5 and 0 for gpt-4o-mini.

    Measured . Source: data/broker-same-model/results.json · the field note.

  2. On a sealed corpus of private changes, reviewed by every model through the same tool-free seat, the cheapest model, openai/gpt-oss-120b, cost $0.000631 per review and caught 29 of 60 seeded defects; the dearest, anthropic/claude-opus-5.5, cost $0.083274 per review and caught 31 of 60. The catch intervals of all models in the experiment overlap, so on this corpus the price of a review varied far more than what it caught.

    Measured . Source: data/review-frontier/results.json · the field note.

What is not a result yet.

  • Whether routing by Odin's own rulebook beats a fixed model on real tasks. EXP 018, “Does routing by our own rulebook beat one fixed model on real tasks?”: its pre-registered verdict is incomplete, so this brief cites none of its figures. The record and the field note say why.
  • Whether the routed arm holds up over fresh runs. EXP 019, “Routing by our own rulebook, measured again: three fresh runs of the routed arm”: Results are published: the record and the field note. This page does not summarise them yet.
  • Whether Odin's per-call cost estimates match what the brokers bill. EXP 015, “Does Odin's cost estimate match what the brokers bill?”: Pre-registered; no result is published yet. Its measurement window: to .
  • That Odin's routing saves money against a fixed-model baseline. No experiment has measured it, and this brief makes no savings claim.
  • How much traffic Odin can route to a provider. Not measured, and not promised here. Any volume would be agreed in writing.
  • Negotiated prices. None exists. Every price the lab has paid so far is a broker's own published price, a promotion or prepaid credit.
  • Provider partnerships. None is signed. A provider is named as a partner only after a signed agreement is on record.

The rule for this page.

No projection, forecast, logo or partner appears here. Each number is read from the experiment record it links to when the site is built, and the build stops if a number is not read from an experiment record or no longer matches it (scripts/number-provenance.mjs). Selected evidence and open questions and how we run experiments hold the rest.

Start a conversation