---
title: "Drex Router: per-turn model routing for agents"
description: "Drex Router picks a model and reasoning effort for each agent turn in about 57 to 87 ms. Call the selected model with your existing keys and gateway."
canonical: "https://www.nace.ai/drex/router"
updated: 2026-11-19
---

# Drex Router: per-turn model routing for agents

> Drex Router picks a model and reasoning effort for each agent turn in about 57 to 87 ms. Call the selected model with your existing keys and gateway.

## Drex Router

A small model that picks the right LLM or SLM and reasoning effort for every step your agent takes. Your agent makes the call.

- 57–87 ms per routing decision
- 1/28 the cost of Jev Router at near-equal quality
- 87.6% on MMLU at $0.10 per 1,000 questions

| Agent | Step | Task | Candidates | Routed to | Effort |
| --- | --- | --- | --- | --- | --- |
| expense-agent | 2 | 15% of €240 | Llama 3.1 8B, GPT-6 Luna, Ministral 3B, Gemini 3.8 Flash, Claude Haiku 5.5, Claude Sonnet 5.5 | Ministral 3B | No reasoning |
| code-agent | 7 | Rename getUser | Qwen small, DeepSeek V4.1 Flash, GPT-6 Luna, Claude Sonnet 5.5, Flint, Claude Opus 5.5 | Flint | Low effort |
| trip-agent | 1 | Plan a Lisbon trip | GPT-4o mini, Claude Opus 5.5, Gemini 3.8 Flash, GPT-6 Luna, GPT-6 Sol, Claude Sonnet 5.5 | Claude Opus 5.5 | High effort |
| ap-agent | 4 | Label invoice lines | gpt-oss-120b, Mistral Small 3.2, GLM 5.3 Flash, Llama 3.1 8B, GPT-6 Luna, Claude Sonnet 5.5 | Llama 3.1 8B | Low effort |
| ci-agent | 3 | Find why tests flake | Claude Sonnet 5.5, Gemma 4 31B, DeepSeek V4.1 Flash, GPT-6 Luna, Claude Haiku 5.5, GPT-6 Sol | Claude Sonnet 5.5 | Medium effort |
| support-agent | 2 | Tag ticket #4821 | Ministral 3B, Gemma 4 31B, GPT-4o mini, Gemini 3.8 Flash, Claude Haiku 5.5, GPT-6 Sol | GPT-4o mini | No reasoning |
| inbox-agent | 5 | Pull the due date | Qwen small, GLM 5.3 Flash, DeepSeek V4.1 Flash, GPT-6 Luna, Claude Opus 5.5, Gemini 3.8 Flash | Gemini 3.8 Flash | Low effort |

## Three questions, three models

| User message | Routed to | Effort | Turn cost |
| --- | --- | --- | --- |
| What’s 15% of €240? | GPT-6 Luna | No reasoning | $0.001 |
| Is Lisbon rainy in March? | Claude Sonnet 5.5 | Low effort | $0.003 |
| Plan 5 days in Lisbon for two, under €1,500. | Claude Opus 5.5 | High effort | $0.050 |

Spent $0.054 against $0.080 with Claude Opus 5.5 on every turn.

Illustrative model spend in USD. All Opus uses Claude Opus 5.5 for every turn.

## Route to any model

17 models are built in. Route to anything else by sending its price, context window and tags. No retraining.

- Large: Claude Opus 5.5, Claude Sonnet 5.5, GPT-6 Sol, GPT-4 1106 preview
- Fast: GPT-6 Luna, Gemini 3.8 Flash, Claude Haiku 5.5, DeepSeek V4.1 Flash, GLM 5.3 Flash, GPT-4o mini
- Open weights: gpt-oss-120b, Gemma 4 31B, Llama 3.1 8B, Mistral Small 3.2 24B, Mistral Small, Mixtral 8x7B, Ministral 3B
- Your profiles: Qwen, Flint by Recursive, Your fine-tune

## Same quality, a fraction of the cost

### Answer quality vs cost

| Router | Judge score (0–10) | Cost per 1,000 prompts |
| --- | ---: | ---: |
| Drex Router | 8.80 | $0.28 |
| Jev Router | 8.84 | $7.80 |
| Jev 1.14 | 8.65 | $2.37 |

306 prompts from the 2026 development sample. Log cost axis.

### MMLU accuracy vs cost

| System | Accuracy | Cost per 1,000 questions |
| --- | ---: | ---: |
| Drex Router | 87.6% | $0.10 |
| Jev Router | 86.4% | $1.09 |
| 12B base | 74.1% | $0.33 |

1,400 MMLU questions, answered end to end. Log cost axis.

## One call before every turn

1. Send the turn: Messages and up to 255 candidates, at every step.
2. Get a decision: The pick and a probability for every candidate.
3. Call the model: With your own keys or gateway.

Request:

```http
POST /v1/route
{
  "messages": [
    { "role": "system",
      "content": "You are trip-agent. Step 1: plan the itinerary." },
    { "role": "user",
      "content": "Plan 5 days in Lisbon, under €1,500" }
  ],
  "candidates": [
    "openai/gpt-6-luna",
    "anthropic/claude-sonnet-5.5",
    "anthropic/claude-opus-5.5",
    { "id": "flint", "provider": "self-hosted",
      "price_in": 0.1, "price_out": 0.4,
      "context": 262144, "tags": ["coding", "small"] }
  ],
  "cost_weight": 0.3
}
```

Response:

```json
{
  "selected": "anthropic/claude-opus-5.5",
  "probabilities": {
    "anthropic/claude-opus-5.5": 0.81,
    "anthropic/claude-sonnet-5.5": 0.15,
    "openai/gpt-6-luna": 0.03,
    "flint": 0.01
  },
  "request_id": "req_8f2c…"
}
```

- [Open Console](https://console.nace.ai/?utm_campaign=drex-router-launch&utm_content=router&utm_medium=website&utm_source=nace.ai)
- [API docs](https://console.nace.ai/docs?utm_campaign=drex-router-launch&utm_content=router&utm_medium=website&utm_source=nace.ai)
- [MCP server](https://console.nace.ai/docs/developer-tools/mcp?utm_campaign=drex-router-launch&utm_content=router&utm_medium=website&utm_source=nace.ai)

## Frequently asked questions

**Does Drex Router proxy my requests?** No. It returns a decision. Your code calls the selected model with your own keys, directly or through your gateway or LiteLLM.

**What model is it built on?** A 12B Gemma-based Nace decision model. It reads conversations of up to 32,768 tokens.

**What happens if routing fails?** The API returns an error and no pick: 504 on a timeout, 503 if routing is unavailable. You aren’t charged, and your agent calls its default model.

**Can I tune it on my own traffic?** Yes. Talk to us about tuning Drex Router on turns from your agent.

[Drex](https://www.nace.ai/drex)

## For agents

- [Drex API guide](https://www.nace.ai/drex.md): the decision model: request shape, question types, benchmarks and deployment, as markdown
- [Decision endpoint](https://console.nace.ai/v1/systemone): POST with model "drex-latest", a state string and typed questions (choice, noul, score); https://console.nace.ai/v1/models lists checkpoints
- [Document endpoints](https://console.nace.ai/v1/documents/{parse|split|classify|extract|ground}): POST a source (a public URL, an uploaded workspace file or a finished parse job) and poll GET https://console.nace.ai/v1/documents/jobs/{id}; the same key and credit as decisions
- [Console](https://console.nace.ai): API keys, the playground, usage and billing; send `Authorization: Bearer $NACE_API_KEY` on every request
- [Console docs](https://console.nace.ai/docs): quickstart, question types, the document routes, errors, limits and pricing
- [Nace Document Intelligence](https://www.nace.ai/perception.md): the Perception series: parse, split, classify, extract and ground any file type with a location on every value
- [Drex for Search](https://www.nace.ai/drex/search.md): Drex for Search: decision-powered retrieval for agents, Drex finds the page, cited search through the CLI, SDKs, MCP and REST, and the public benchmark ranks
- [Industries](https://www.nace.ai/industries.md): Nace by industry: the files Document Intelligence reads and the decisions Drex makes for accounting, banking, mortgage, insurance, legal, healthcare, government, construction, logistics and professional services
- [Developers](https://www.nace.ai/developers.md): The execution layer for agents: Perception, Decision and Search behind one key, the Drex Max frontier, install steps, a scripted agent session and the four deployment boundaries
- [Site index](https://www.nace.ai/llms.txt): every page on nace.ai; append `.md` to any page URL or send `Accept: text/markdown` for the markdown version
