Model routing
Everything the router may reach, who may reach it, and the ordered rules that decide. Rules are evaluated top-down; the first match wins, and every rule declares a fallback chain so a degraded provider or an exhausted budget never turns into a failed request.
Models
Prices are USD per 1M tokens. Blended is the 75/25 input/output mix used across this demo.
| Model | Provider | State | Rank | In | Out | Blended | Access | Enabled |
|---|---|---|---|---|---|---|---|---|
| local | ollama | ga | 1 | $0.00 | $0.00 | $0.00 | all | |
| together-llama | together | ga | 2 | $0.20 | $0.20 | $0.20 | all | |
| claude-haiku | anthropic | ga | 3 | $0.25 | $1.25 | $0.50 | all | |
| bedrock-haikuadded live | bedrock | shadow | 3 | $0.80 | $4.00 | $1.60 | all | |
| bedrock-sonnet | bedrock | ga | 5 | $3.00 | $15.00 | $6.00 | all | |
| claude-sonnet | anthropic | ga | 5 | $3.00 | $15.00 | $6.00 | all | |
| claude-opus | anthropic | ga | 8 | $15.00 | $75.00 | $30.00 | production, dev-* |
Default model per team
This is the team's default when no routing rule matches ahead of it. Changes are recorded in the configuration audit log.
Routing order and fallbacks
Evaluated top-down. Fallbacks fire on the listed conditions, in order, before the request is ever failed.
production incidents, outages, security escalations, customer-facing breakage — correctness dominates cost
writing, reviewing or debugging code; architecture; APIs; infrastructure
chores and non-engineering work — formatting, renames, dependency bumps, doc touch-ups, general knowledge, drafting, research
anything not matched above
How real prompts resolve
The same three probes used in the demo, replayed against the rules above.
| Prompt | Rule | Model | Note |
|---|---|---|---|
| our checkout is returning 500s since the last deploy | incident-response | claude-opus | |
| draft a launch email for the Q3 release | routine-and-non-dev | local | marketing is over budget; rule 3's target is free, so no fallback needed |
| review this migration for lock contention | code-and-review | bedrock-sonnet | availability fallback 1 of 2 · equivalent rank 5 · +0% cost |
Configuration audit log
Every enable, disable and assignment change in this session, with before and after values.
Aggregate catalog spend to date: $1473.00 across 7 models.