LLM compute · available
One key. Every LLM. The price that fits the task.
LLMs differ less in capability than their price lists suggest — and far more in price. We route every request to where the measured performance is good enough for its purpose, at the lowest price of the day and within the commitments you set.
How it works
The technology behind it, without the marketing fog.
The API speaks the widely used OpenAI format: chat messages, streaming via server-sent events, tool calls, structured outputs. Existing code changes the base URL and the key — nothing else.
Region, zero data retention, quantization level: whatever you commit to excludes providers before any comparison happens. A cheaper route that breaks a commitment simply does not exist for the selection.
You choose a tier per use case — coding, knowledge, instruction following. Which model answers behind it is decided by measurement, not by a vendor's marketing.
Input, output and cached input are counted separately and billed in CHF. Without credit, a request is rejected before it reaches the model — not after the answer.
The research field: inference economics and model routing
The running costs of an LLM increasingly arise at inference, not in training. Three levers determine them: the size of the active model per token (mixture-of-experts activates only a fraction of the parameters), the number format (from 16 bit through FP8 down to 4-bit formats with microscaling) and the reuse of computation (prefix and KV caching, speculative decoding).
Routing goes one step further: instead of sending every request to the strongest model, a router estimates which model will solve it well enough. Cascades try small first and escalate only when uncertain. The difficulty lies in the prediction: the quality of an answer can only be measured after it exists, while the router has to decide beforehand.
Open questions
- How well can the difficulty of a request be predicted from its text?
- When does an escalation cost more than it gains in quality?
- How stable is the measured ranking of models across weeks and provider updates?
If the breakthrough succeeds
Outlook, not a promise: what would become possible once the open questions are answered.
Predictable routing
If the required model tier could be reliably predicted per request, most everyday requests would be answered by small models. Worked example: if a model at a tenth of the price answers four out of five requests equally well, costs fall by about two thirds — at unchanged quality.
Quality commitments instead of model names
A customer would no longer book a model but an outcome: “code that passes the tests, at this price”. Switching providers would become invisible and automatic.
Where we stand
The same evidence as on the overview — every number with its kind.
- OpenAI-compatible API, one key for all platforms
- Quality tiers per use case instead of a model lottery (model taylor)
- Live price comparison across platforms, card payment only
Who it is for
- Product teams that want to use several models without several contracts
- Agencies and developers with fluctuating load
- Companies with requirements on region and data retention
Usable today.
Get a key, swap the base URL, start — billed per token in Swiss francs.
Go to LLM Broker