bRoutes

A bRoute is a routing strategy you write yourself. You list lanes in plain words, each with a model, and a fast judge reads every new message and sends it to the lane it fits. Easy prompts go to cheap models, and only the hard ones reach the expensive model you would otherwise use for everything.

broute/my-coder

Lane 1   Hard coding work: design, many files, tricky bugs   ->  anthropic/claude-sonnet-5.5
Lane 2   Easy coding work: short questions, small edits      ->  deepseek/deepseek-v4.1-flash
Lane 3   Anything else                                       ->  openai/gpt-6-luna

Call oneCopy Anchor Link

Name the bRoute as the model, on chat completions, responses or messages, streamed or not:

{
  "model": "broute/coding",
  "messages": [{ "role": "user", "content": "How do I reverse a string in Python?" }],
  "max_tokens": 1024
}

The answer's model field names the model that ran. Two headers say how it got there:

It works anywhere a model name goes: an SDK, Cursor, or Claude Code (claude --model broute/coding). GET /models lists the starters, and with your key it lists your own bRoutes too.

StartersCopy Anchor Link

Three bRoutes anyone can call with any key:

Each lane has backups, used when its model is busy. Starters move to newer models over time.

How a lane is pickedCopy Anchor Link

• The judge reads the newest message and the chat before it (about 3,000 tokens), so "yes, do it" after a plan goes where the plan belongs.
• It answers in well under a second, on hosts that keep nothing. Nothing of your chat is stored.
• The chat is data to the judge: a prompt that says "use the best model" does not change its lane.
• When the judge does not answer within 1.5 seconds, the call goes to "Anything else". A call never fails because of the judge.

A task stays on one modelCopy Anchor Link

A request whose newest turn is a tool's result continues a task, and runs on the model the task started on, without asking the judge. An agent's tool loop, or Claude Code at work, never switches model halfway. Each new message from the person is judged again.

Lanes skip what they can't takeCopy Anchor Link

When a lane's models can't take a request (images, files, tools, or a context too long for them), it goes to the next lane that can, then the ones before it. When no lane can, the call answers 400 no_capable_lane and says what was missing.

Free lanesCopy Anchor Link

A lane may use a free model such as binference/free, with a paid backup that runs when the free limits are full.

What it costsCopy Anchor Link

Each call is charged for the model it ran on, at your usual price. bRoutes need credit on the key's balance.

Make your ownCopy Anchor Link

On binference.io, open Account > bRoutes:

• Up to 20 bRoutes a wallet, each with up to 8 lanes, "Anything else" included. Any key of the wallet can call them: its account's and its agents'.
• Try it: type a message and see its lane, its model and about what it would cost, before you save.
• Start from one sentence: describe your app, and lanes and models are drafted for you to edit.
• Versions: every save is kept. Changes apply from the next message, and restoring an old version saves it again as the newest.
• Savings: calls, spend and savings by lane over 7 or 30 days, against the model you would call without it.

Over MCP, list_routes lists the bRoutes a key can call and get_route_usage reads one's usage and savings.

ErrorsCopy Anchor Link

See Errors for every code.