Which LLM is the cheapest for our workload?

The best-for pages rank the live catalogue per task by the index that measures it, and the router picks the cheapest model that meets your quality rule for each request.

How much can we save on our LLM bill?

On the public benchmark the router saved 89% against sending everything to the reference model. Import last month's provider bill on the bill page and the estimate is computed on your own traffic before you change a line of code.

How often do model prices change?

The catalogue is refreshed from the vendors' price lists; each model page shows the date of its data.

Open source

An open-source LLM router with zero dependencies. Run it where your data lives.

Apache-2.0, a Python package with no runtime dependencies, a container image and a Helm chart. Route between models, agents and tools behind your own firewall with policy, a trace per request and learning from outcomes - or use the hosted platform on a free plan.

Get a free API key Route this request in the playground

Free plan, no card. Open source under Apache-2.0; self-hosting is free forever.

From the benchmark

runtime dependencies in the Python package

The core installs nothing else; YAML, the server and OpenTelemetry are optional extras. Every release is on PyPI and npm with its checksum, and the images carry the same version.

The whole benchmark, with the cases the router lost

The problem

Routing your traffic through someone else's service is a data decision, not only a cost one.

A hosted router sees every prompt. For regulated data, air-gapped networks or a procurement process that asks where the bytes go, the router has to run where the models run - without a second product, a second licence or a dependency tree to audit.

Prompts must not leave the network
A dependency with its own dependencies is a security review
The hosted and the self-hosted version must be the same software

How the router answers it

Three things it does for every request.

A library first
pip install opensmartroute pulls in nothing else. The router is a class; targets and rules are two YAML files; the trace is a Python object.
A service when you want one
osr serve exposes /route, /v1/chat/completions and an MCP server from the same catalogue; the container image and the Helm chart run it in your cluster.
The platform, self-hosted
The hosted platform - workspaces, keys, dashboard, savings ledger - ships as the same open-source package with a Compose bundle. Enterprise features are a setting, not a second product.

Watch one decision

Illustrative: the lowest cost that still clears the quality the task needs. The live quote below is the real one.

A real request

A catalogue with an on-premises model and two cloud models, and a rule that pins personal data at home.

The same two files drive the library, the server, the container and the MCP server.

Route it in the playground and read the trace

Exampleyaml

# rules.yaml
- id: pii-stays-home
  when: { contains_pii: true }
  prefer: [onprem-llama]
  pin: true          # not a preference - a rule the score cannot override

$ osr -t targets.yaml -r rules.yaml route "Summarise the attached patient letter"
target=onprem-llama  confidence=0.91  policy: cloud targets rejected (pii)

Where this stands next to the alternatives, and how to go deeper.

Install

docker run --rm -p 8000:8000 crosrua2c6onyidguw.azurecr.io/opensmartroute:latest

Package, image, Compose bundle and Helm chart

Or no install at all: the hosted API answers a plain curl with the decision, and the browser extension shows the price and the router's pick under the composer of the AI chat sites you already use.

For the business

What it means for which models you pay for.

Catalogue of every vendor
The right model, not the famous one
Price, context, abilities and benchmark indexes side by side, with the model that fits each task named.
The rankings
89% less
Lower cost per request
On the 150-prompt benchmark the router spent 89% less than sending every prompt to the reference model, and the judged quality went up, not down.
Read the benchmark
Per million tokens, every model
Prices you can compare
The catalogue prints input and output prices and what three typical monthly workloads cost on each.
The model catalogue

Which LLM is the cheapest for our workload?: The best-for pages rank the live catalogue per task by the index that measures it, and the router picks the cheapest model that meets your quality rule for each request.
How much can we save on our LLM bill?: On the public benchmark the router saved 89% against sending everything to the reference model. Import last month's provider bill on the bill page and the estimate is computed on your own traffic before you change a line of code.
How often do model prices change?: The catalogue is refreshed from the vendors' price lists; each model page shows the date of its data.

Route your first request in under a minute.

OpenSmartRoute is open source and the free plan keeps the full trace of every decision. Create a workspace, mint a key and send the request above.

Create a free account Read the benchmark

Free plan, no card. Fifteen thousand decisions a month with the full trace.

# rules.yaml - id: pii-stays-home when: { contains_pii: true } prefer: [onprem-llama] pin: true # not a preference - a rule the score cannot override $ osr -t targets.yaml -r rules.yaml route "Summarise the attached patient letter" target=onprem-llama confidence=0.91 policy: cloud targets rejected (pii)

An open-source LLM router with zero dependencies. Run it where your data lives.

Routing your traffic through someone else's service is a data decision, not only a cost one.

Three things it does for every request.

A library first

A service when you want one

The platform, self-hosted

A catalogue with an on-premises model and two cloud models, and a rule that pins personal data at home.

Where this stands next to the alternatives, and how to go deeper.

What it means for which models you pay for.

The right model, not the famous one

Lower cost per request

Prices you can compare

Route your first request in under a minute.

An open-source LLM router with zero dependencies. Run it where your data lives.

Routing your traffic through someone else's service is a data decision, not only a cost one.

Three things it does for every request.

A library first

A service when you want one

The platform, self-hosted

A catalogue with an on-premises model and two cloud models, and a rule that pins personal data at home.

Where this stands next to the alternatives, and how to go deeper.

What it means for which models you pay for.

The right model, not the famous one

Lower cost per request

Prices you can compare

Route your first request in under a minute.