AI Systems · Architecture

Model Routing and Fallback Systems

Model routing with an LLM router: cheap versus strong pins, canaries, eval-based routes, version pins, and when a fallback model should take over a request.

Read time
15 min
Level
Intermediate
Updated
2026-08-25
Components
Request classifier · Model catalog · Canary lane
The short answer

A model routing system is an LLM router that sends each request to a cheap, fast, or strong model based on task class and eval scores, pins versions, canaries changes, and fails over along a declared fallback chain when a provider errors, times out, or fails a quality gate.

Key takeaways

  • 01

    Route by task difficulty and risk, not by a single default frontier model; extraction and classification rarely need the expensive tier.

  • 02

    Pin model versions. Unpinned "latest" is an unscheduled deploy of someone else's weights.

  • 03

    Fallbacks are a chain with the same schema contract — not a silent swap to a model that cannot call your tools.

  • 04

    Canary a percentage of traffic against the golden set before promoting a new pin; eval-based routing beats gut feel.

  • 05

    If every request is high-stakes or the catalog has one model, you do not need a router — you need a pin and a health check.

01

Model routing: cheap, strong, and fallback

Model routing is the control plane that chooses which model handles a call. An LLM router classifies the request (or the agent step), looks up a catalog of pinned models with cost, latency, context, and tool-support flags, and dispatches. A fallback model is the next catalog entry when the primary errors, exceeds SLO, or fails a cheap quality check.

The point is not cleverness. It is to spend frontier tokens on planning, judgment, and hard generation, and to spend small-model tokens on extraction, classification, rewriting, and judges. Declarative routes — task type → model pin → fallback list — are testable. A prompt that says "pick a model" is not.

  • Classifier or explicit task-type from the orchestrator
  • Catalog: pin, context window, tools, cost, region
  • Canary lane with eval comparison
  • Fallback chain with schema and tool compatibility
02

The catalog, pins, and compatibility

The catalog is the source of truth: provider, model ID, exact version or snapshot, region, max tokens, structured-output support, tool-calling support, and data-handling terms. Agents and routers reference aliases ("planner-strong", "extract-fast") that resolve to pins. Changing a pin is a deploy.

Compatibility matters more than raw quality. A fallback that cannot emit your JSON schema or cannot call tools is not a fallback; it is a different product. Record capability flags and refuse to route a tool-using span to a completion-only model.

03

Eval-based routing and canaries

Choose routes with evidence. Offline, run the golden set on each candidate pin and keep a route only if it meets the quality bar for that stratum at a cost and latency you accept. Online, canary a few percent of traffic to a candidate pin, compare judge scores and hard graders, and promote or roll back.

Some routers score difficulty live (small classifier, or a cheap model that may escalate). Escalation must be bounded: one upgrade per request is typical. Open-ended "keep trying bigger models" is how a ten-cent request becomes a dollar without a better answer.

04

Fallback chains and health

A fallback chain is ordered: pin A → pin B (same capability class) → degrade (cached answer, extractive snippet, or a typed error). Trigger on timeouts, 5xx, rate limits, and optionally on schema-invalid output after retries. Do not fallback on 4xx that indicate a bad request — that loops.

Health of a provider is an input to the router, not a surprise. If error rate on pin A crosses a threshold, shed new traffic to pin B and page. User-visible traces should show which pin served the request so support is not debugging the wrong model.

05

Failure modes and when not to route

Classic failures: unpinned aliases that change overnight; a fallback that drops tool support; a classifier that sends hard legal questions to the cheap tier; canaries without a rollback; routers that retry four models and bill all of them. Cost dashboards that ignore fallbacks under-count.

Skip a router when you have one qualified model, or when every request is in the same high-risk class (then pin the strong model and add health-check failover only). Also skip live difficulty classifiers you have not eval'd — a misrouted 5% of traffic is a quality incident that looks like a model problem.

  • Unpinned "latest" versions
  • Incapable fallbacks (no tools, no schema)
  • Unbounded escalation and retry billing
  • Classifier errors on high-risk strata
06

Operating the router

Treat route tables like code: review, CI against the golden set, canary, then promote. Dashboard cost per task class, fallback rate, and per-pin quality. Pin versions on a calendar so provider deprecations are planned. Work stays model-agnostic — Anthropic, OpenAI, Google, Mistral, open-source, or a client fine-tune — because the catalog, not the app, names the weights.

Frequently asked questions

What is an LLM router?

An LLM router is the control plane that assigns each request or agent step to a pinned model from a catalog, based on task class, cost, latency, and capabilities such as tools or JSON. It canaries new pins and fails over along a declared chain. It is configuration plus health, not a prompt that "picks a model."

What is model routing in production?

Production model routing sends easy work to cheap, fast models and hard or high-risk work to stronger pins, with eval evidence for each route. Versions are pinned, changes are canaried, and traces record which pin served the call. Unpinned aliases are treated as unscheduled deploys.

What is a fallback model?

A fallback model is the next compatible pin in a chain when the primary times out, errors, or fails a cheap quality gate. Compatible means it can honor the same schema and tools. A weaker model that cannot call functions is a degraded mode, not a fallback, and should be labeled as such.

Should you pin LLM versions?

Yes. Pin the exact version or snapshot in the catalog and promote new pins through evals and a canary. "Latest" lets a provider change behavior without your CI. Deprecations then become a planned catalog change rather than a Monday incident.

When do you not need model routing?

When one qualified model handles the product, or when every request is the same high-risk class. In those cases pin that model, add health-check failover to a compatible twin if you must, and skip live classifiers. A router without evals is just another way to mis-spend tokens.

Keep reading

ArchitectureMulti-Agent Orchestration SystemsQualityLLM Evaluation Systems (Evals)TrainingRL Environments for Agent TrainingArchitectureAgent Memory SystemsArchitectureAgentic RAG SystemsArchitectureComputer-Use Agent SystemsArchitectureFunction-Calling and Tool-Use SystemsArchitectureGraphRAG SystemsArchitectureHuman-in-the-Loop AI SystemsArchitectureHybrid Retrieval and Re-ranking SystemsQualityLLM Guardrail SystemsInfrastructureLLM Inference and Serving SystemsInfrastructureLLM Observability and TracingInfrastructureMCP Tool Gateway SystemsTrainingPEFT and Fine-Tuning PipelinesArchitecturePermissioned Retrieval SystemsQualityPrompt Injection Defense SystemsArchitectureRAG ArchitectureArchitectureRealtime Voice AI SystemsQualityAI Red-Teaming SystemsArchitectureStructured Generation SystemsTrainingSynthetic Data Generation Systems

Building one of these systems?

We help teams design, build, and validate production AI systems — orchestration, evals, and training environments included.

FAQ

Working with us

How soon can AI systems work start?

Typically within a week or two of a scope being agreed. The first delivery is deliberately a small batch so you can check the output against your expectations before volume ramps.

What do you need from our team?

One process owner who knows the workflow, one engineer with access to the systems involved, and a weekly 45-minute review. No standing committee, and no requirement for an ML specialist on your side.

Who owns the output and the data?

You do. Datasets, labels, weights, evaluation suites and runbooks are yours and are handed over at the end. Your data trains your models only, with zero-retention provider settings by default.

Can you scale volume up quickly if we need it?

Yes, and the quality bar holds because the rubric and gold set are already agreed by that point. Ramping is a staffing question, not a re-scoping one, so it usually takes days rather than a new engagement.

We already have a vendor for this. Why switch?

Often you should not. The cases where teams move to us are when they cannot get a quality number out of their current vendor, or when the work is delivered as an opaque batch with no trace of how disagreements were resolved.

What happens after the engagement ends?

We stay on-call for 30 days at no extra cost, then move to an optional support retainer. Most teams also keep a quarterly evaluation review with us to catch drift early.

Copyright © 2026
ReinforcedX, Inc.
All rights reserved