The router
Free
It's free and open source: MIT licence, no account, no hosted service in between. That does not change with Pro.
Source on GitHub →LLM Autorouter · open source (MIT)
An open-source router on your own machine. Our small routing model, s1-llm-auto-router, reads every prompt, and the router sends it to the cheapest model in your list that is likely to answer it well. OpenAI-compatible: change the base URL and model name.
The router is free and open source. Pro is not available yet: its button leads to a waitlist.
export OPENROUTER_API_KEY=sk-or-... # replace with your own OpenRouter key
python3 -m venv ~/.llm-autorouter && . ~/.llm-autorouter/bin/activate
pip install "auto-model-router[s1-router] @ git+https://github.com/fstandhartinger/auto-model-router@80a09591e4da97c5ee141c976d10b4afbefe36aa"
curl -fsSLo router.yaml https://decisionmodels.io/examples/llm-autorouter.yaml
AUTO_ROUTER_CONFIG=router.yaml auto-model-routerInstall
The commands above take your OpenRouter key, create a virtual environment, install the router with the routing model, fetch a starter config with three models on OpenRouter and start the router. The routing model (183 MB) downloads from Hugging Face on the first request and is checked against a pinned SHA-256 before it runs.
python3-venv; on macOS a current Python (for example brew install python); on Windows use WSL.Point any OpenAI-compatible client at http://127.0.0.1:8787/v1 with the model name auto. The response headers x-router-model, x-router-category and x-router-reason say which model answered and why.
curl http://127.0.0.1:8787/v1/chat/completions \
-H "Content-Type: application/json" \
-d '{"model":"auto","messages":[{"role":"user","content":"What is the capital of France?"}]}'Other providers, local models, Claude Code as a gateway and the full configuration: README on GitHub.
How it works
The router runs on your machine as an OpenAI-compatible endpoint. Your provider keys stay in your environment.
One CPU pass answers seven questions about the request: topic, difficulty, what a wrong answer would cost, whether it needs tools, an image or a long context, and whether it continues the previous turn. Only a scrubbed, truncated summary of the turn is classified.
For each model in your list it weighs the price of the call (including the prompt cache), the chance the model gets this kind of task right and the cost of a retry, using per-model prices and benchmark data from Benchmark Heaven (which draws on Artificial Analysis and Epoch AI). Then it forwards the request with your key.
What it costs
The router
It's free and open source: MIT licence, no account, no hosted service in between. That does not change with Pro.
Source on GitHub →Routing model on your CPU
s1-llm-auto-router is open weights under the MIT licence on Hugging Face. The starter config above runs it locally (backend: local-s1); your prompts are not sent to us.
Routing model over our API
Pay per use at the launch price (USD 0.001), output tokens free. Set backend: decisionmodels and your DECISIONMODELS_API_KEY; nothing runs on your CPU. Prices excl. applicable taxes.
The models that answer your prompts are billed by their providers on your own keys, for example OpenRouter. The router estimates the cost of each request; it does not bill you for anything.
Pro · planned
Planned price: one week free, then EUR 10 per installation and month, plus the per-use price of s1-llm-auto-router over our API. Nothing can be bought or used today, and we have not set a start date. What we plan for Pro:
The free router keeps every routing feature it has today.
Measured
Our own measurements only. The samples are small; each line says what it was measured on.
Routing model · held-out test set
On 2,175 held-out synthetic routing requests, s1-llm-auto-router named the right category 89.8 % of the time and led to the same model as the reference labels in 74.2 % of cases. That set is synthetic and mostly labelled by the same LLM-ensemble process as the training data. On 100 separately, blind-labelled requests it scored 88 % and 60 % (about ±8 points at that size). Read the limits on the model card.
Model card →Routing model · speed
Median per decision on one or two threads of a 12-core x86-64 server once loaded; loading takes 2 to 3 s at the first request, peak memory 650 MB. The model card measured 26 ms on four dedicated threads of an older Xeon; small or busy CPUs are slower.
Router README →Router · 40 tasks, 29 Sep 2026
40 synthetic, public-safe coding, maths, extraction and HTML tasks, each run once through the router and once on Claude Opus 5.5, first answer only (no answer check, no retry): mean model cost USD 0.000337 against USD 0.006412 per task; 30 against 29 tasks passed, a difference this small sample cannot separate from chance (95 % interval −7.5 to +12.5 points). On the ten coding tasks the router passed 7, Opus 9. The router answered 7.7 s slower on average. This run used the router's earlier classifier (Weiche-395M), not s1-llm-auto-router.
Method and data →Router · 20 Claude Code tasks, 26 Sep 2026
20 coding tasks in Claude Code, once with Claude Opus 5.5 for every turn and once through the router (settings of 26 Sep, earlier classifier) with its answer check on, which sends rejected answers to a stronger model: USD 3.21 against USD 7.73 at list prices, but hidden tests passed 172 of 179 against 177 of 179, and both blind judges preferred the Opus run on 13 of 20 tasks. With the answer check off it was 13× cheaper but passed only 117 of 179. The starter config above has no answer check.
Results in the README →Install check, 15 Nov 2026: we ran the install in a fresh Python 3.12 virtual environment on a Linux server. Our test prompts were classified by s1-llm-auto-router on the CPU; a one-word question and a coding task both went to GLM-5.3 Flash, the cheapest model in the starter config.
FAQ
With the starter config, s1-llm-auto-router runs on your CPU and none of your prompts are sent to Decision Models. Your prompt goes to the model the router picks, on your own key. Without your prompts, the router downloads the routing model from Hugging Face once and fetches public model prices and benchmark data from benchmarkheaven.com. With the hosted routing model, a scrubbed and truncated summary of each turn is also sent to api.decisionmodels.io.
Claude Code can use the router as its gateway, and Claude Code, Codex, opencode and Cursor can hand sub-tasks to it through its MCP tool. Anthropic does not support routing Claude Code to non-Claude models and does not allow passing a Claude.ai login on for others; the README describes each setup and the vendors’ terms.
No. Pro does not exist yet. The router and the routing model are free to run; only the hosted routing model is billed, per use.