Modern .NET applications increasingly use more than one AI model.
A single application may have access to:
A high-capability model for complex reasoning
A faster model for simple requests
A lower-cost model for routine workloads
A specialized model for a particular task
A fallback model for availability problems
The engineering challenge is deciding which model should handle each request.
A static model-selection strategy is straightforward:
Application
|
v
Selected Model
|
v
Response
A routing strategy adds another decision layer:
Application
|
v
Routing Layer
|
+---- Model A
+---- Model B
+---- Model C
|
v
Response
The advantage is flexibility, but routing also introduces additional decision logic and potentially additional latency.
Microsoft.Extensions.AI provides the IChatClient abstraction and composable chat-client pipelines, which makes it possible to place routing, logging, retry, configuration, and other behaviors around AI clients. The current API also provides ChatClientBuilder and DelegatingChatClient as mechanisms for composing these pipelines.
This article explains how to benchmark a routing-based chat client against static model selection and determine whether routing actually improves the application's overall performance and economics.
Introduction
Suppose an application supports three request categories:
Simple Question
|
v
Fast Model
Complex Reasoning
|
v
Advanced Model
Fallback
|
v
Backup Model
A static strategy might send every request to the same model:
Every Request
|
v
Model A
A routing strategy might inspect the request and select a model:
+--> Model A
|
Request --> Router --+--> Model B
|
+--> Model C
The routing strategy can potentially improve cost, latency, or availability.
However, the router itself has a cost.
It may introduce:
Therefore, routing should be treated as an engineering hypothesis that needs to be benchmarked.
What Is Static Model Selection?
Static model selection means the application chooses the model before processing the request and does not dynamically change that choice.
For example:
IChatClient client = primaryClient;
var response = await client.GetResponseAsync(
"Explain dependency injection in .NET.",
cancellationToken: cancellationToken);
The application knows exactly which client will process the request.
This approach is simple and predictable.
The execution path is:
Request
|
v
Static Selection
|
v
Model
|
v
Response
What Is Routing?
Routing introduces a decision mechanism between the application and the underlying model clients.
Request
|
v
Router
|
+--> Fast Model
|
+--> Capable Model
|
+--> Fallback Model
The router can use different strategies.
For example:
Request Complexity
|
+--> Simple ----> Fast Model
|
+--> Moderate -> Balanced Model
|
+--> Complex --> Advanced Model
Another strategy could use availability:
Primary Model
|
X
Unavailable
|
v
Fallback Model
A third strategy could combine both:
Complexity
+
Cost
+
Availability
+
Latency
|
v
Model Selection
Microsoft.Extensions.AI and IChatClient
The IChatClient abstraction provides a common interface for chat model interactions. This allows application code to work against an abstraction instead of depending directly on a specific model implementation.
This is important for benchmarking because the application can execute the same workload against different client configurations.
For example:
public interface IModelExecutor
{
Task<ChatResponse> ExecuteAsync(
string prompt,
CancellationToken cancellationToken);
}
A static implementation can wrap one model.
A routing implementation can select among several models.
The benchmark can then compare both using the same test scenarios.
Static Selection Architecture
A simple static architecture looks like this:
Application
|
v
IChatClient
|
v
Model A
The advantage is that there is almost no selection overhead.
The main limitation is that every request follows the same model path unless application code explicitly changes the client.
Routing Architecture
A routing architecture looks like this:
Application
|
v
Routing Layer
|
+---------+---------+
| | |
v v v
Model A Model B Model C
The router becomes responsible for determining the target.
This can be implemented as a custom IChatClient wrapper or as a component in a chat-client pipeline.
DelegatingChatClient is specifically designed as a base type for clients that wrap another IChatClient, and the chat-client pipeline can be composed using ChatClientBuilder.Use(...).
Define the Benchmark Question
Before measuring anything, define what the benchmark is trying to prove.
For example:
Does dynamic routing reduce cost without causing unacceptable latency or quality degradation compared with static model selection?
That question produces several measurable dimensions:
Latency
Cost
Quality
Success Rate
Fallback Rate
Routing Accuracy
Throughput
Without a clear hypothesis, it is easy to produce a benchmark that generates numbers without providing an engineering conclusion.
Benchmark Scenarios
Use multiple workload categories.
Simple Requests
Examples:
What is dependency injection?
Convert this JSON into a C# record.
What does HTTP 404 mean?
Complex Requests
Examples:
Analyze this architecture and identify scalability risks.
Explain the tradeoffs between two distributed-system designs.
Review this code and identify concurrency problems.
Long-Context Requests
These contain larger amounts of input and can expose different model behavior.
Failure Scenarios
Simulate:
Timeout
Rate Limit
Unavailable Model
Invalid Response
Transient Network Failure
Routing should be evaluated not only when everything works but also when the preferred model fails.
Establish a Baseline
The first benchmark should use static selection.
For example:
All Requests
|
v
Model A
Measure:
p50 latency
p95 latency
p99 latency
Token usage
Cost
Success rate
Quality score
This becomes the baseline.
Then execute the same workload through the router.
All Requests
|
v
Router
|
+--> Model A
+--> Model B
+--> Model C
The two measurements can then be compared.
Benchmark Harness
Create a common interface.
public interface IBenchmarkClient
{
string Name { get; }
Task<BenchmarkResponse> ExecuteAsync(
BenchmarkRequest request,
CancellationToken cancellationToken);
}
The request can contain:
public sealed record BenchmarkRequest(
string Id,
string Category,
string Prompt);
The response can contain:
public sealed record BenchmarkResponse(
string RequestId,
string Model,
TimeSpan Latency,
long InputTokens,
long OutputTokens,
bool Success);
This provides a consistent measurement format.
Measure Latency
Use a monotonic timer.
var start = Stopwatch.GetTimestamp();
var response = await client.ExecuteAsync(
request,
cancellationToken);
var elapsed =
Stopwatch.GetElapsedTime(start);
This measures application-observed execution time.
Do not include unrelated operations such as loading the benchmark dataset or writing the final report inside the timed region.
Measure Routing Overhead Separately
Routing latency should not be hidden.
Consider:
Total Routed Latency
=
Routing Decision
+
Model Request
+
Response Processing
If routing itself requires another model call:
Total Latency
=
Router Model Call
+
Target Model Call
That additional call can be significant.
If routing is rule-based:
Total Latency
=
Rule Evaluation
+
Target Model Call
The difference can be substantial.
Therefore, capture routing time independently:
var routingStart = Stopwatch.GetTimestamp();
var target = await router.SelectAsync(
request,
cancellationToken);
var routingLatency =
Stopwatch.GetElapsedTime(routingStart);
Then measure the actual model request separately.
Benchmark Static Selection
A static benchmark might look like:
public async Task<BenchmarkResponse> RunStaticAsync(
BenchmarkRequest request,
IChatClient client,
CancellationToken cancellationToken)
{
var start = Stopwatch.GetTimestamp();
var response = await client.GetResponseAsync(
request.Prompt,
cancellationToken: cancellationToken);
var latency =
Stopwatch.GetElapsedTime(start);
return new BenchmarkResponse(
request.Id,
"static-model",
latency,
GetInputTokens(response),
GetOutputTokens(response),
true);
}
The exact token-usage extraction depends on the provider and client implementation.
The benchmark should use the actual usage metadata available from the selected client rather than estimating token counts from string length.
Benchmark Routing
A routing benchmark follows the same measurement boundary:
public async Task<BenchmarkResponse> RunRoutedAsync(
BenchmarkRequest request,
IRoutingClient client,
CancellationToken cancellationToken)
{
var start = Stopwatch.GetTimestamp();
var response = await client.GetResponseAsync(
request.Prompt,
cancellationToken);
var latency =
Stopwatch.GetElapsedTime(start);
return new BenchmarkResponse(
request.Id,
response.Model,
latency,
response.InputTokens,
response.OutputTokens,
true);
}
The important point is that both strategies receive the same benchmark request.
Rule-Based Routing
The simplest routing approach uses deterministic rules.
public string SelectModel(BenchmarkRequest request)
{
return request.Category switch
{
"Simple" => "fast",
"Complex" => "advanced",
"LongContext" => "long-context",
_ => "fast"
};
}
This has almost no classification overhead.
It is also easy to test.
The disadvantage is that rules can become increasingly complicated as workloads grow.
LLM-Based Routing
A more dynamic strategy can use a model to classify the request.
User Request
|
v
Routing Model
|
+--> Simple
+--> Complex
+--> Specialized
|
v
Target Model
This can be flexible but introduces an additional model operation.
For example:
Routing Decision = 80 ms
Target Model = 600 ms
Total = 680 ms
If static selection requires only:
Target Model = 600 ms
the router has made the request slower.
Routing must therefore generate enough savings elsewhere to justify its own overhead.
Benchmark Routing Accuracy
A routing system should also be evaluated for decision quality.
Create an expected model category for each benchmark request:
public sealed record RoutingExpectation(
string RequestId,
string ExpectedRoute);
Then compare:
Expected Route
vs
Selected Route
Calculate:
Routing Accuracy =
Correct Decisions
-----------------
Total Decisions
A router that selects the wrong model frequently may not produce the expected cost or quality benefits.
Measure Model Quality
Latency alone is not sufficient.
Suppose:
Static Model
Latency: 800 ms
Quality: 0.95
Routing
Latency: 650 ms
Quality: 0.86
Routing is faster, but the quality regression may be unacceptable.
Depending on the application, measure:
Task success rate
Structured-output validity
Groundedness
Answer relevance
Domain-specific correctness
Human evaluation
Retrieval quality for RAG workloads
The exact evaluation metric should match the application.
Cost Measurement
For each model request, capture:
Input Tokens
Output Tokens
Model
Pricing Version
Calculated Cost
Then aggregate by route.
Static Strategy
--------------
Total Cost
Average Cost
Cost Per Successful Task
Routing Strategy
----------------
Total Cost
Average Cost
Cost Per Successful Task
The most useful comparison is often:
Cost Per Successful Task
rather than simply cost per API request.
Example Cost Comparison
Imagine a benchmark produces:
| Metric | Static Selection | Routing |
|---|
| Requests | 1,000 | 1,000 |
| Success Rate | 96% | 97% |
| p50 Latency | Measure | Measure |
| p95 Latency | Measure | Measure |
| Total Tokens | Measure | Measure |
| Total Cost | Measure | Measure |
| Cost / Successful Task | Measure | Measure |
The benchmark should populate these values from actual measurements.
Avoid inserting illustrative numbers into a production recommendation unless they come from a reproducible test.
Fallback Routing
Routing can also be used for resilience.
Primary Model
|
X
Failure
|
v
Fallback Model
A benchmark should measure:
Primary Success Rate
Fallback Rate
Fallback Latency
Final Success Rate
For example:
Request
|
v
Primary
|
X
|
v
Fallback
|
v
Response
The latency of the failed primary attempt should remain visible.
Otherwise, the fallback benchmark may appear faster than it actually is.
Routing and Retries
Retries introduce another variable.
Suppose the router sends a request to Model A.
Model A
|
X
Retry
|
X
Fallback Model
The final request may succeed, but the total cost includes both failed attempts.
Track:
Attempts
Models Used
Retry Count
Total Latency
Total Cost
Final Status
This provides a much more accurate picture of routing behavior.
Warm-Up Strategy
Do not use the first request as the only benchmark measurement.
Warm up each client before collecting steady-state measurements.
foreach (var client in clients)
{
await client.ExecuteAsync(
warmupRequest,
cancellationToken);
}
Then start collecting measurements.
Run cold-start tests separately if cold-start behavior matters to the production workload.
Run the Same Query Set
The static and routed systems should receive exactly the same workload.
Benchmark Dataset
|
+------> Static
|
+------> Router
Do not allow the router to receive easier questions than the static system.
A fixed dataset also makes regression testing easier.
Query Distribution Matters
Suppose the real application receives:
70% Simple
20% Moderate
10% Complex
but the benchmark contains:
20% Simple
30% Moderate
50% Complex
The resulting cost and latency numbers may not represent production.
Use a representative distribution.
If several workloads are important, benchmark them separately and report the results independently.
Concurrency Testing
A routing strategy can behave differently under load.
Test:
1 concurrent request
5 concurrent requests
10 concurrent requests
25 concurrent requests
50 concurrent requests
depending on service limits and the target workload.
Measure:
p50
p95
p99
Throughput
Error Rate
Route Distribution
A router that performs well at one request at a time may behave differently when multiple requests compete for the same model capacity.
Route Distribution
Record how frequently each model is selected.
For example:
Model A: 60%
Model B: 30%
Model C: 10%
This is important for cost analysis.
If the router unexpectedly sends 80% of requests to the expensive model, the expected savings may disappear.
Static vs Routing Comparison
A useful comparison table is:
| Dimension | Static Selection | Routing |
|---|
| Implementation complexity | Low | Medium/High |
| Selection overhead | Minimal | Depends on strategy |
| Model flexibility | Low | High |
| Cost optimization | Limited | Potentially strong |
| Failover | Explicit application logic | Can be centralized |
| Debugging | Simple | More complex |
| Observability requirements | Moderate | Higher |
| Workload adaptation | Limited | Stronger |
| Predictability | High | Depends on routing policy |
Routing is not automatically better.
It is better when the additional complexity produces measurable value.
Common Benchmarking Mistakes
Comparing Different Prompts
The workload must remain consistent.
Ignoring Router Latency
A routing decision is part of the request path.
Measuring Only Average Latency
Always examine tail latency.
Ignoring Routing Accuracy
A poor route can increase cost or reduce quality.
Using Only Simple Queries
Routing benefits often appear when workloads have meaningful variation.
Ignoring Failure Paths
Fallback behavior should be benchmarked explicitly.
Ignoring Cost of Retries
A successful fallback may still have incurred multiple failed model calls.
Comparing Different Model Configurations
Keep relevant settings consistent where the benchmark is intended to isolate routing behavior.
Treating Quality as Secondary
A cheaper or faster response is not necessarily a better response.
Observability
A routing system should record enough telemetry to explain every decision.
Useful attributes include:
TraceId
RequestId
SelectedModel
RoutingReason
RoutingLatency
ModelLatency
InputTokens
OutputTokens
RetryCount
FallbackUsed
EstimatedCost
Success
This allows engineers to answer:
Why did this request use Model B?
How long did routing take?
How much did the request cost?
Did fallback occur?
Was the response successful?
These questions become essential when debugging production behavior.
Building a Routing Wrapper
Because DelegatingChatClient is designed for wrapping an inner IChatClient, a custom routing abstraction can follow the same compositional pattern. The important design choice is to keep routing policy separate from model execution.
A simplified conceptual implementation could look like:
public sealed class RoutingChatClient
{
private readonly IReadOnlyDictionary<string, IChatClient> _clients;
private readonly IRoutingPolicy _policy;
public RoutingChatClient(
IReadOnlyDictionary<string, IChatClient> clients,
IRoutingPolicy policy)
{
_clients = clients;
_policy = policy;
}
public async Task<ChatResponse> GetResponseAsync(
string prompt,
CancellationToken cancellationToken)
{
var route = await _policy.SelectAsync(
prompt,
cancellationToken);
var client = _clients[route.Model];
return await client.GetResponseAsync(
prompt,
cancellationToken: cancellationToken);
}
}
This is a simplified example rather than a complete implementation of a production routing client.
In a real application, the routing layer should also handle:
Cancellation
Resilience
Telemetry
Model availability
Policy validation
Error classification
Fallback
Cost tracking
Composing the Client Pipeline
The ChatClientBuilder API supports composing intermediate chat-client stages. This allows routing-related behavior to coexist with logging, retries, options configuration, function invocation, and other middleware-like components.
A conceptual pipeline can look like:
Application
|
v
Routing
|
v
Logging
|
v
Retry
|
v
Model Client
The exact ordering should be chosen deliberately.
For example, placing telemetry around the routing layer can help measure routing decisions separately from downstream model latency.
Release Regression Testing
Once the benchmark works, run it automatically.
Code Change
|
v
Build
|
v
Benchmark Dataset
|
+--> Static Baseline
|
+--> Routing Strategy
|
v
Compare
|
+--> Latency
+--> Cost
+--> Quality
+--> Reliability
|
v
Release Decision
For example:
Routing must satisfy:
p95 latency <= baseline + 20%
Quality >= baseline - 5%
Cost per successful task < baseline
Error rate <= baseline
The exact thresholds should be based on application requirements.
When Static Selection Is Better
Static model selection is often preferable when:
The workload is highly predictable.
One model already satisfies quality requirements.
Routing logic does not produce meaningful savings.
Simplicity is a major requirement.
The additional routing latency is unacceptable.
There are few model options.
A simpler architecture can be the better architecture.
When Routing Is Better
Routing becomes more attractive when:
Requests vary significantly in complexity.
Different models have different strengths.
Cost optimization is important.
Availability requirements justify fallback paths.
The application handles multiple workload classes.
The routing decision can be made reliably.
The operational team can observe and debug the routing behavior.
Advantages
Better Model Utilization
Different workloads can use different models.
Potential Cost Reduction
Simple requests can avoid unnecessarily expensive models.
Better Resilience
Fallback routing can improve availability when a preferred model fails.
Centralized Policy
Model-selection rules can be managed in one place.
Easier Model Evolution
New models can be introduced without rewriting every application workflow.
Disadvantages
Additional Complexity
Routing adds another component to the request path.
Routing Latency
A model-based router can add another AI operation.
Debugging Complexity
A response can depend on both the routing decision and the selected model.
More Telemetry
Engineers need visibility into route decisions, fallback, retries, and model usage.
Potential Quality Regression
An incorrect route can select a model that is cheaper or faster but less capable for the task.
Best Practices
Establish a static-model baseline before evaluating routing.
Use the same benchmark dataset for both strategies.
Measure routing latency independently.
Track p50, p95, and p99 latency.
Measure routing accuracy.
Track route distribution.
Measure token consumption and cost.
Include model quality in the benchmark.
Test fallback and retry behavior separately.
Run concurrency tests.
Use realistic production query distributions.
Keep model configuration consistent during controlled comparisons.
Record routing decisions in telemetry.
Compare cost per successful task rather than raw request cost alone.
Automate benchmark execution as part of regression testing.
Frequently Asked Questions
Is RoutingChatClient always better than static model selection?
No. Routing introduces additional complexity and potentially additional latency. It should be used when dynamic model selection provides measurable value.
Does routing always reduce AI costs?
No. A router can increase costs if it adds another model call, selects expensive models too frequently, or causes additional retries.
Should routing be rule-based or AI-based?
Start with deterministic rules when they are sufficient. AI-based classification can provide more flexibility, but it introduces additional latency and evaluation complexity.
What should I measure when benchmarking routing?
At minimum, measure latency, cost, quality, routing accuracy, success rate, fallback rate, token usage, and route distribution.
Should the router itself be included in latency?
Yes. If the goal is to measure user-visible request latency, the routing decision is part of the request path and should be included in total latency. It should also be measured separately so its overhead is visible.
How can I prove that routing is worthwhile?
Compare routing with a static baseline using the same workload and evaluate whether it produces an acceptable improvement in cost, latency, reliability, or quality after accounting for routing overhead.
Conclusion
Dynamic model routing is an attractive architecture for applications that work with multiple AI models, but it should not be adopted simply because multiple models are available.
The right question is whether routing produces measurable value compared with a well-defined static baseline.
A useful benchmark evaluates the complete picture:
Routing Overhead
+
Model Latency
+
Cost
+
Quality
+
Reliability
+
Fallback Behavior
A routing strategy may reduce cost for simple workloads, improve resilience during model failures, and select more capable models for complex requests. At the same time, it can introduce classification latency, additional operational complexity, and incorrect model selections.
The most reliable approach is therefore empirical: establish a static baseline, run the same workload through the routing strategy, measure p50/p95/p99 latency, cost, quality, route accuracy, and failure behavior, and then make the architecture decision from those results.
In production AI systems, model routing should be treated as a measurable optimization layer rather than an assumption that dynamic selection is automatically better.