AgentNuro Control · Model Routing

Model routing: send each request to the right model

AgentNuro routes each request to the model that fits on cost, quality, and latency. Let the platform decide, write your own rules, or run both. Every route shows the dollars it saved.

Try Control Lite, free See how routing decides

app.cloudnuro.ai/routingRequest routed

715 tokens, summarize, difficulty 0.45

gemini-2.5-flash0.776 quality, $0.000419

gpt-4o-mini0.775 quality, $0.000510

o4-mini0.796 quality, $0.001100

Routed by AgentNuro

$0.000419

vs ~~$0.005624~~ on opus-4.1

Saved $0.005205, 92.5% lower

Not what you would pick? Run this request on your own choice, any time.

What is model routing?

Model routing directs each AI request to the most cost-effective model that still meets the quality bar, downgrading where it is safe and keeping premium models where they matter.

Without right-sized routing

Overkill by default.Simple requests go to the most expensive model.
Uncontrolled token consumption.Nobody checks whether a cheaper model would have cleared the same bar.
No fit check.Provider dashboards show spend by model, but not whether each request needed that model.

Spend by request type

This month, total AI spend

$41,208

RequestDefaultRight-sizedSaved

Classify support ticketopus-4.1gemini-2.5-flash$0.0041

Translate release notesopus-4.1gemini-2.5-flash$0.0038

Draft follow-up emailopus-4.1gpt-4o-mini$0.0046

Review contract for risk clausesopus-4.1opus-4.1kept

Draft board strategy memoopus-4.1opus-4.1kept

Saved is per request against the default model. Every request defaults to opus-4.1 whether or not the task needs it, until it is right-sized.

How does routing decide which model to use?

Where you have set a rule, your rule applies. For everything else, each request passes through four stages before it runs. Switch the scenario to see what happens when the best-scoring model is not approved for your organization.

gemini-2.5-flashIn your model catalog, Support team

The top pick is approved, so it runs as chosen. Switch it off to see the fallback.The top pick is not approved. Stage 3 swaps in the closest approved fit.

1
Prompt Predictor
Reads the prompt and predicts token count, request type, and difficulty, then shortlists candidate models.
This request715 tokens, summarize, difficulty 0.45, 3 candidates
2
Model Selector
Scores every candidate on quality, cost, and latency and picks the best overall combination.
Top pickgemini-2.5-flash wins at 0.776 quality for $0.000419, against o4-mini at 0.796 for $0.001100
3
Approved-model check
Confirms the pick is approved in your organization’s catalog. If it is not, the closest approved fit is used instead.
Catalog check
gemini-2.5-flash

gemini-2.5-flashgpt-4o-mini

gemini-2.5-flash is approved for this team. No substitution.gemini-2.5-flash is not approved for this team. Closest approved fit: gpt-4o-mini.
4
AI Provider
Runs the request on the selected model through the gateway, and logs the decision for audit.
ResultRouted at $0.000419, versus $0.005624 on opus-4.1. Fully logged.Routed at $0.000510, versus $0.005624 on opus-4.1. Fully logged. Saved $0.005205, 92.5% lowerSaved $0.005114, 90.9% lower

Sample data, for illustration only. This sequence runs inside AgentNuro Control on every request.

Who decides how requests are routed?

You do. Let AgentNuro decide, write your own rules, or run both. Nothing reroutes outside the models you have approved, and you can change your mind at any time.

Default

AgentNuro decides

Every request is scored on quality, cost, and latency and sent to the best fit among your approved models. Nothing to configure, and the dollars saved show up on every route.

Your rules

You write the rules

Match on task complexity, token count, request type, agent, intent, or project budget. Send matches to a model, to cache, or to the cheapest approved model, with a fallback if it is unavailable.

Always

You keep the final say

Approve or restrict models in your catalog, switch any rule off in one click, shadow-test a rule on a share of traffic first, or override the choice on a single request.

Run both

Rules you set apply to the workloads you name. Everything else follows AgentNuro’s cost-and-quality determination, so you stay in control of what matters without writing a rule for everything.

How much does each route save?

Every route carries its own saving, whether AgentNuro chose it or one of your rules did. You see the dollars next to the decision, so keeping, changing, or switching off a route is never a guess.

app.cloudnuro.ai/routing/routesSample figures

AgentNuro scores every request on quality, cost, and latency and picks the best fit among your approved models. No rules to write.

RequestWho decidesRouted toSaved per request

Classify a support ticketAgentNuro decidescost, quality, latencygemini-2.5-flash$0.0041

Summarize a sales callAgentNuro decidescost, quality, latencyhaiku-4.5$0.0038

Draft a follow-up emailAgentNuro decidescost, quality, latencygpt-4o-mini$0.0046

Review a contractAgentNuro decidescost, quality, latencysonnet-4.5$0.0021

Every route showsDollars savedQuality changeFallback if unavailableOn or off in one click

Illustration with sample figures, not a product screen. Saved is per request, against the model the application originally called.

Does routing affect quality?

Quality is part of every routing decision, and every decision can be overridden.

Judged on more than price

Every candidate is scored on quality, cost, and latency together, so the cheapest model never wins by default.

Near identical in practice

In the worked example above, the winner scored 0.776 on quality against 0.796 for the top-scoring candidate, at less than half the cost.

Never locked in

A recommendation, not a mandate. Override it on any request, switch a rule off, or restrict a model in your catalog.

Quality delta, month to date

−0.4%

Inside the 2% tolerance

0%2% tolerance

Sample data, for illustration only.

Quality shown per rule
Each rule reports its own quality change next to the dollars it saves, so a saving never hides a regression.
Shadow tests
Try a rule on a small share of traffic first. Its projected saving is shown but kept out of the total until you rely on it.

How is AgentNuro’s routing different?

Most routing tools make you pick a side: write every rule yourself, or hand the decision to a black box. AgentNuro lets you do either, and shows the money either way.

Rules-only gatewaysAutomatic-only routers

AgentNuro Control

Who decides

Rules-only gatewaysYou, through conditions you write and maintain.Automatic-only routersThe router, through its own model of cost and quality.

AgentNuro ControlYou choose: the platform’s determination, your rules, or both.

Getting started

Rules-only gatewaysConfigure conditions, often as JSON or a flow, before anything routes.Automatic-only routersPoint calls at the router and tune it with cost and quality settings.

AgentNuro ControlA one-line endpoint change starts automatic routing. Add rules when you want them.

Dollars saved

Rules-only gatewaysTypically cost analytics. Comparing against a baseline is left to you.Automatic-only routersTypically a benchmark claim or a pre-sales estimate.

AgentNuro ControlSaved per request, per rule, and month to date, routing and cache shown separately.

Quality guardrail

Rules-only gatewaysWhatever your conditions encode.Automatic-only routersA learned prediction of which model will answer best.

AgentNuro ControlA quality tolerance, with the quality change reported on every rule.

Rolling out a change

Rules-only gatewaysPercentage splits or versioned routes, on some gateways.Automatic-only routersTypically retrained or reconfigured by the vendor.

AgentNuro ControlShadow-test a rule on a share of traffic. Its projected saving stays out of the total.

Allowed models

Rules-only gatewaysWhatever each route lists.Automatic-only routersThe set of models you pass in.

AgentNuro ControlA catalog with Approved, Restricted, and Evaluating status. Routing stays inside it.

The first two columns describe common approaches across the AI gateway and router market, not a specific vendor.

Frequently asked questions

A broad range of model providers through AgentNuro Control, including OpenAI, Anthropic (Claude), Google Gemini and Vertex AI, AWS Bedrock, and Azure AI Foundry. Seat-based tools like Microsoft 365 Copilot are attributed by licensed user rather than routed.

Yes. Rules can match on task complexity, token count, request type, agent, intent, similarity to earlier prompts, regional health, and project budget. Each rule sends matching requests to a model, to cache, or to the cheapest approved model, with a fallback and an on/off switch. You can shadow-test a rule on a share of traffic before it counts. Anything outside your rules follows AgentNuro’s cost and quality determination.

Yes. By default AgentNuro scores every request on quality, cost, and latency and sends it to the best fit among the models you have approved. There is nothing to configure, and the dollars saved still show up on every route.

Nothing is forced. Accept the recommendation, or run the request on your own choice, any time, per request. You can also restrict a model in your catalog or switch a rule off.

Each routed request is compared with what it would have cost on the model your application originally called. Savings appear per request, per rule, and month to date, with routing and cache savings shown separately. A shadow test shows a projected saving that is not counted in the total.

Token count, request type, and difficulty from the prompt itself, then quality, cost, and latency across the shortlisted candidates. Whether the winner is approved in your organization is checked last, and the closest approved fit is used if it is not.

One line. Routing runs in the AgentNuro gateway, so your calls go through it with a one-line endpoint change, or through the SDK.

Control fails open. If the gateway can’t be reached, requests fall back to calling the provider directly using the last known routing rule, so an outage pauses routing, not your application.

Routing reads each prompt in flight to estimate its size, type, and difficulty, which is how it picks a model. Prompt content is not stored. What is logged is the decision: the scores, the model chosen, any substitution, and the cost.

Stop overpaying for the model everyone defaults to.

A one-line endpoint change starts automatic routing, and every route shows what it saved against the model your app originally called.

Try Control Lite, free