Cheapest or fastest: routing a model without pinning it
A gateway can run one model on many upstream providers. Tarsk lets you ask for the cheapest or the fastest endpoint per model, and keeps fallbacks open.

A model name does not point at one machine. OpenRouter runs the same model on several upstream providers, and those providers charge different prices and answer at different speeds. Tarsk used to leave that choice to the gateway. Now you can ask for the cheapest or the fastest route on any model, per model, from the composer.
The choice you could not make
You pick a model for what it can do. Behind it, the gateway balances requests across providers that differ in price, throughput and precision. A cheap endpoint can serve a compressed version of the model. A fast endpoint can be a different company’s hardware on a given day.
Sometimes you want a specific one of those. A long agent run with many tool calls is easier to reason about when the endpoint stays quick. A batch of short drafts costs less on a cheaper route. Before this, neither request had anywhere to go, and the gateway’s default applied to every turn.
Where the choice lives
Open the model menu in the composer and hover a row. The row normally shows a relative price. On hover, a brain icon takes that spot. Press the brain and a popover opens with two groups.
- Reasoning sets how much thinking the model does, on models that support levels.
- Routing sets Auto, Cheapest or Fastest, on models whose provider offers a route choice.
The popover floats above the menu, so scrolling the model list does not clip it. Your pick is stored with that model’s settings. Every chat that uses the model gets it, and you can set a different route on a different model.
Routing appears for OpenRouter models today. Other providers decide their own endpoint, so the group stays hidden there rather than offering a switch that does nothing.
What each choice asks for
Auto is the default. Tarsk sends no routing instruction and the gateway chooses, exactly as before. Use it unless you have a reason not to.
Cheapest asks the gateway to sort providers by price. The tooltip warns that a cheaper endpoint may be quantized and answer worse. Compression saves money and can cost you precision on hard questions.
Fastest asks for the highest throughput. The provider may change between turns, because the sort happens on each request rather than once per conversation.
Each choice also tells the gateway to skip endpoints that cannot carry the request. Tools, cache controls and reasoning settings must survive the route. Without that guard, a sorted request could land on an endpoint that drops a tool definition.
A preference, not a pin
Tarsk never locks a model to one endpoint. It sorts the available providers and keeps fallbacks allowed, so a provider that goes down or rejects the request does not end your turn.
The other option was to pin a single endpoint. A pin gives you a predictable provider until that provider has a bad minute, and then you own a failure that the gateway would have routed around. A pin also breaks prompt caching more often, because each provider caches its own prefix.
The trade-offs worth knowing
Routing can change an answer without changing anything you see. A compressed endpoint may produce a weaker reply on the same prompt, and the chat looks identical. That is the cost of a stated preference. The alternative was a hidden default.
Sorting also turns off the gateway’s load balancing, so a run can move to a different provider in the middle of a conversation. That provider starts with a cold cache for your prompt. Tarsk reports the cold cache in its prompt cache diagnostics instead of pinning the first provider to protect it.
The setting lives with the model, so it applies wherever that model runs, including chats on a Remote Machine and on iOS. The iPhone client has no routing controls; it uses the value you set on the desktop.
Pick a route, then forget it
Three habits cover most work:
- Leave Auto on models you trust the gateway to place.
- Set Cheapest on models you use for short, low-stakes drafts.
- Set Fastest on the model that drives long tool-heavy runs.
Open the model menu, hover the model, and press the brain. The Routing group is there.