Skip to content
NuRoute

Artemis picks the model

Send model: "auto". Artemis chooses by expected quality and cost, and the NuRoute API runs the request.

Try ArtemisRead the benchmarks

Pick the trade-off

  • Cost-first

    98.1% cheaper than gpt-5.5 · quality 0.905

  • Balanceddefault

    97.5% cheaper than gpt-5.5 · quality 0.952

  • Quality-first

    72.1% cheaper than gpt-5.5 · quality 0.967

Measured on 426 held-out prompts.

Same quality, a fraction of the cost

  • Artemis · Cost-first0.905

    $104 per 100,000 requests · quality (0 to 1)

  • DeepSeek-V4-Flash0.943

    $130 per 100,000 requests · quality (0 to 1)

  • Artemis · Balanced0.952

    $141 per 100,000 requests · quality (0 to 1)

  • Artemis · Quality-first0.967

    $1,558 per 100,000 requests · quality (0 to 1)

  • Claude Sonnet 50.949

    $3,988 per 100,000 requests · quality (0 to 1)

  • gpt-5.50.977

    $5,596 per 100,000 requests · quality (0 to 1)

  • Balanced matched Claude Sonnet 5’s quality (0.952 against 0.949) at 3.5% of its cost.
  • A single cheap model comes close at Balanced: DeepSeek-V4-Flash for every request scored 0.943 at $130 per 100,000 requests, against Artemis’s 0.952 at $141.

426 held-out prompts, our own test. Tables, conditions and limits

How it decides

  1. 1

    Read the task

    A small encoder, not an LLM, scores the prompt across 24 task types.

  2. 2

    Score every model

    Expected quality from graded answers. Expected cost from the predicted answer length.

  3. 3

    Pick by preset

    Best expected quality after a penalty for extra cost. The preset sets the penalty.

Three presets

Artemis presets and their measured held-out results against gpt-5.5
PresetCost savedQuality keptPer 100k requests
Cost-first98.1%92.7%$104
Balanced97.5%97.5%$141
Quality-first72.1%99.1%$1,558

Measured on held-out prompts, against gpt-5.5 answering every request. Each preset also has a fitted target, which is not a guarantee.

The API

One endpoint. The response headers tell you what Artemis decided.

Request
curl https://your-gateway/v1/chat/completions \
  -H "Authorization: Bearer aicp-..." \
  -d '{ "model": "auto",
        "messages": [{ "role": "user", "content": "Extract the invoice number: ..." }] }'
Response headers
X-Router-Version: v2
X-Router-Capability-Version: <capability version>
X-Router-Model: <the model that served the request>
X-Router-Estimated-Saving: <estimated fraction saved>

your-gateway is your NuRoute address; angle-bracket values are placeholders. Full reference

Good to know

  • Structured output (response_format) is ignored for now.
  • Multi-turn: each request is decided on its system prompt and latest message.
  • Tool calling is checked at the provider level, not for model reliability.
  • Artemis chooses only among its pool. To use another model, name it.
  • Not in self-hosted NuRoute yet.

Questions

Does Artemis run my request, or only recommend a model?

It runs it. Artemis is the model-selection step in the NuRoute API. There is no recommendation-only endpoint today.

Which API keys does it use?

A NuRoute API key to call it, and either NuRoute’s providers or your own provider keys. See authentication.

Does routing add another model call?

No. Two small models run the decision. No extra generative call and no tokens.

Are the quality targets guaranteed?

No. Targets are fitted on our training prompts. Test a preset on your own traffic before relying on it.

Need budgets, policies and fallback too?

Explore Gateway