Guide · Architecture
A hybrid router for high-stakes agentic work
Begin with a narrow, reviewable split: route multi-file agentic coding and high-ambiguity tasks to the higher-capability lane; route stable classification, extraction, and straightforward drafting to the lower-cost lane.
Evidence key: Derived — the editorial source publishes one date for the whole response, not a date per article.
Source: ModelScale guide /articles/hybrid-router (reviewed 2026-09-17). Provider links in the reference should be checked before purchase.
1. Decide whether a hybrid route is justified
When is a two-model route operationally safer and economically better than sending every request to one endpoint?
2. State the assumptions that make the route auditable
A blended result is only useful when the route can be reconstructed later. Record the conditions that determine dispatch before comparing projected cost.
- Traffic can be labeled by task type before dispatch.
- Quality loss has a measurable business cost and an escalation mechanism.
- Latency and price are measured under the same host and workload conditions.
- Fallback behavior is tested independently from the primary route.
3. Separate facts, derived relationships, and interpretation
Evidence quality changes what a routing result can support. Keep editorial judgment visibly separate from observed prices, service measurements, and production outcomes.
| Statement | Type | What to verify |
|---|---|---|
| Request mix changes the blended monthly cost. | Derived relationship | Input/output mix and host prices |
| A high-capability lane is worth its cost for complex work. | Interpretation | Acceptance rate, recovery cost, human review |
What to verify: Input/output mix and host prices
What to verify: Acceptance rate, recovery cost, human review
4. Compare routing policies on the same cost basis
Compare routing policies on one cost basis rather than on impressions. Take the request mix from your own telemetry — how many requests each lane actually served, and the input and output tokens each consumed — then price that same mix against current published rates for every candidate model.
- Record the share of requests each lane served over a representative period, not a single day.
- Record input and output tokens separately, because output is priced higher and a routing change moves the two independently.
- Price the same mix for each policy you are considering, so the only variable is the routing decision.
- Re-run the comparison when prices change rather than carrying an earlier result forward.
Current per-model prices for every model in the catalog are on /models, and /subscribe-vs-api compares a subscription against metered API use for the same workload. Use those rather than a figure captured earlier, and treat any cost difference as one input alongside acceptance rate, recovery cost and review effort.
5. Choose the route and its guardrail together
The routing condition, preferred lane, and operational guardrail form one decision. Separating them makes a cheap path look safer than it is.
| Condition | Preferred route | Guardrail |
|---|---|---|
| Multi-file change, unclear acceptance criteria | Capability lane | Human review on sampled completions |
| Stable extraction, bounded output | Economy lane | Schema validation and fallback |
| Latency breach or provider incident | Fallback lane | Circuit breaker and event log |
Guardrail: Human review on sampled completions
Guardrail: Schema validation and fallback
Guardrail: Circuit breaker and event log
6. Continue with the decision surfaces
Use the framework to build a shortlist, inspect model constraints, and compare candidates under the same assumptions.
Models workbench Inspect price, performance, and lifecycle evidence. Make it yours Re-rank models around the six capability weights. Compare models Carry two to four candidates into a detailed trade-off view.