The router in detail

What decides where your prompt goes

The least expensive model is rarely the right answer for every prompt. These are the judgements leap makes on each turn, and all of them are visible to you afterwards.

Your prompt is read before anything is spent

How hard the task looks, whether code or an image is involved, how much context an answer needs, and what this conversation has already covered. A brief follow up after a long piece of work inherits that history instead of being taken at face value.

Capability is settled before cost

Anything unable to hold your document, see your image or call the tool you asked for drops out first. No saving is worth a model that cannot do the task.

A ceiling buys the strongest option under it

Where nothing within your limit clears the quality bar, leap takes the most capable model you can afford rather than the cheapest available, and says plainly that it cannot vouch for the result.

A newer release does not automatically win

One version only retires another at matching size and tier. A small recent model will not displace the larger predecessor it was never built to replace.

Repeated context is charged less

Vendors ask for less when material has been seen moments earlier, and each behaves differently. leap follows those rules per provider and counts the saving in your favour, never as grounds for sending demanding work somewhere weaker.

Every model is rated the same way

Each model carries a Leap score, and a single formula produces all of them from independent evaluations under open licences and the results labs publish themselves. The router weighs that rating against what the task needs and what it costs.

The catalogue refreshes continuously as labs ship and withdraw them, so a release from this morning is a candidate this afternoon. Every decision above is recorded per turn on your machine: which model answered, why the rest were ruled out, and what it cost.

Leap scores are based in part on data from Epoch AI (Epoch Capabilities Index and benchmark runs) and OpenRouter (benchmark leaderboard), both licensed under CC BY 4.0. Leap combines and rescales it; neither source endorses leap.