Model Allowlist
Org allowlist set by Platform Admin. Team Admins may narrow — never expand. Agents and the interview reference a Model Class, never a concrete model.
Model Classes
Agents and the interview ask for a Model Class. The Model Dispatcher resolves it to a concrete model here — the one place model identity is set. Swapping the model behind a class is a one-line change; agent IAM regenerates from it.
Proposal authoring, compliance review, and reasoning-heavy agents
Ticket triage, routing, and high-volume tasks
Allowlist
What each class is permitted to resolve to. Toggle a model off to narrow the team — narrowing is guarded when deployed Agents depend on it.
| Model | Family | Context | $/1M in | $/1M out | Resolves | Org | Team |
|---|---|---|---|---|---|---|---|
| claude-opus-4-7 | Claude | 200k | $15.00 | $75.00 | — | allowed | |
| claude-sonnet-4-6 | Claude | 200k | $3.00 | $15.00 | smart | allowed | |
| claude-haiku-4-5 | Claude | 200k | $1.00 | $5.00 | fast | allowed | |
| amazon-nova-pro-v1 | Nova | 300k | $0.80 | $3.20 | — | allowed | |
| meta-llama-3-1-70b | Llama | 128k | $2.65 | $3.50 | — | denied | |
| amazon-titan-text-premier | Titan | 32k | $0.50 | $1.50 | — | denied |
The same pattern runs through tools, skills, and models: the org admin sets the ceiling, team admins shape the floor. A Team Admin can never enable a model the Platform Admin denied — preserving the org-wide model governance posture under FedRAMP CM-7 (least functionality).