Read the grades free. Pay when you act on them.
Every endpoint serving the same open-weight model, graded every night against a published reference, beside its price.
- The full public table, every endpoint named
- Public JSON at
/v1, rewritten nightly - A nightly route list per model: the endpoints to allow in your OpenRouter calls, ready to paste
- MCP server with 10 free tools, including
get_route_listandcheck_before_route - Last night's changes at
/v1/changes.json - Graded by Bitdepth badge for endpoints inside the floor
- An API key: the full row for any endpoint we grade, as often as you like, with no per-call payment
- An email the morning an endpoint you watch changes grade state, declared precision, price or listing
- A status email every Monday on everything you watch
- Up to 200 endpoints, any model we grade
- Add or drop endpoints any time
- Cancel any time from your receipt
- One endpoint's complete record: per-night scores, per-family self-agreement with pair counts, self-check nights
- No key: an x402 client pays and retries on its own
- Settled in USDC before the row is served
- Unlimited endpoints watched
- Monthly dated attestation of your provider mix
- Verified by Bitdepth badge for your customers
- Your rows privately, before each public table
- The grade and drift feed inside your product
- Models you route added to the nightly run
- Direct support
10 free tools. 1 paid per call. Watch on top.
Every tool runs on the MCP server at https://bitdepth.dev/api/mcp/mcp with no key. Every answer carries the date, the method link and the reference it is measured against.
The endpoints to allow for one model, cheapest first, and the ones left out with the reason. Paste the block into your OpenRouter request so you stop getting routed to endpoints that miss the reference.
Should my agent send calls to this endpoint tonight? Returns route true, false or inconclusive, the reason, and the best value alternative.
The cheapest endpoints for a model that stayed inside the noise floor for 7 nights or more, ranked by blended price.
Per model: endpoints listed, how many inside and outside the floor, the price range, and the cheapest endpoint inside it.
Every endpoint for one model: declared precision, nights, scores, price in and out, and state.
One endpoint's row with its night-by-night series.
The precision an endpoint declares on the router, and whether that declaration changed.
Last night's changes in state, declared precision, price and listing. The feed Watch emails come from.
Endpoints whose score moved beyond the noise threshold across the window.
The models graded nightly, with the date and endpoint count.
One endpoint's full record: per-night scores, per-family self-agreement with pair counts, self-check nights.
The morning email when an endpoint you watch changes. Quiet otherwise.
Run the full prompt set against any endpoint now, graded against last night's reference. Including your own route.
A wrapper for the OpenRouter SDK that pins your calls to endpoints inside the floor, on your own key.
The dated monthly record behind an attestation, as JSON.
How Watch works.
What counts as an endpoint?
One provider serving one model at one declared precision, for example deepseek/deepseek-v4-pro:atlas-cloud/fp8. The tags are in the public table, in each model's JSON, and in get_scorecard.
How do I pick them?
Tick them on the Watch page, from the list of every endpoint we grade. The price follows the count. Reply to any Watch email to change the list.
What is a change?
The endpoint moves inside or outside the noise floor, its declared precision changes, its listed price moves, or it appears or leaves the router. The rules are on the method page.
What emails do I get?
A welcome email within the hour, with where each endpoint stands. A status email every Monday on all of them. And an alert the morning after any of them changes, usually early US Eastern.
Numbers against a stated reference. Not a claim about intent.