NeuroRouterSign in

Pay frontier prices only for frontier problems.

Axon, our open-weights router, sends each request to a small model or a frontier model in about 30 ms. One OpenAI-compatible endpoint, and no request logs.

How a request is routed

Most production traffic is extraction, classification and short rewrites. Small models handle those at a fraction of the cost. Axon catches them before they reach a frontier model.

  1. Score

    Axon reads the prompt, system message and tool definitions and outputs a difficulty score. About 30 ms, before any tokens are generated.

  2. Route

    The request goes to the cheapest model expected to clear your quality threshold, set per route. Ticket classification lands on a small model. A race condition goes to a frontier model.

  3. Escalate

    If the small model’s answer comes back with low confidence, the request is retried on a stronger model. The response tells you which model answered.

No request logs

A router sees every prompt, so ours keeps none of them. Send customer records or internal code without leaving a copy with us.

Request logs
None kept
Realtime prompts and outputs
Never stored
Async and batch jobs
Held encrypted until you collect the result, then deleted
Training on your data
Never
What we keep
Token totals per model, for billing

Details are in our Privacy Policy.

Open weights. Run it yourself or let us host it.

Run Axon locally with Ollama in front of any provider you already use, or call it through NeuroRouter. If your quality bar differs from ours, fine-tune it on your own traffic.

Parameters
421M
Download
846 MB
Decision time
About 30 ms
Runs on
Ollama on your hardware, or hosted by us
Release
Weights and training code at public launch

Change two lines: the base URL and the model name. res.model tells you which model answered.

import OpenAI from "openai";

const client = new OpenAI({
  baseURL: "https://api.neurorouter.ai/v1",
  apiKey: process.env.NEUROROUTER_API_KEY,
});

const res = await client.chat.completions.create({
  model: "axon/auto",
  messages: [
    { role: "user", content: "Extract every date from this invoice." },
  ],
});

res.model; // "deepseek-ai/deepseek-v4.1-flash", picked by Axon

To skip routing, call a model by ID, such as moonshotai/kimi-k3.

Models and pricing

19 open models, by name or through Axon. Realtime for interactive traffic, async when minutes are fine, batch for jobs that can wait up to 24 hours.

Interactive traffic.

Model prices in US dollars per million tokens, Realtime tier
ModelTypeCost classInputCached inputOutput
Kimi K3.1 ProMoonshot AIFrontierChatFrontier$2.00$0.20$10.00
DeepSeek V4.1 Flash 1017DeepSeekSmallChatSmall$0.15$0.01$0.60
Muse Spark OpenMetaFrontierChatFrontier$1.25$0.15$4.25
DeepSeek V4.1 FlashDeepSeekSmallChatSmall$0.15$0.01$0.60
GLM 5.3 FlashZ.aiSmallChatSmall$0.15$0.03$0.50
GLM 5.3Z.aiFrontierChatFrontier$1.40$0.28$4.40
Qwen3.8 27BQwenFrontierChatFrontier$0.45$0.04$3.00
DeepSeek V4 Pro 0813DeepSeekFrontierChatFrontier$1.32$0.13$3.96
Muse Glimmer 30BMetaSmallChatSmall$0.11$0.01$0.32
DeepSeek V4 Flash 0731DeepSeekSmallChatSmall$0.09$0.02$0.18
Kimi K3Moonshot AIFrontierChatFrontier$3.00$0.30$15.00
GLM 5.2Z.aiFrontierChatFrontier$0.93$0.19$3.00
Nemotron 3 Ultra 550B A55BNVIDIAFrontierChatFrontier$0.50$0.13$2.20
Qwen3.6 35B A3BQwenSmallChatSmall$0.14$0.14$1.00
Gemma 4 31B ITGoogleSmallChatSmall$0.12$0.12$0.35
Qwen3.5 35B A3B dottxtQwenSmallChatSmall$0.35$0.20$1.20
LightOnOCR 2 1B bbox soupLightOnSmallOCRSmallAsync and batch only
Qwen3 VL 235B A22B InstructQwenSmallVisionSmall$0.21$0.12$1.90
Qwen3 Embedding 8BQwenSmallEmbeddingsSmall$0.04$0.04Not billed

USD per million tokens. axon/auto requests are billed at the rate of the model that answered.

NeuroRouter is in private beta.

Invites go out in small batches. Sign in with the address your invite was sent to.