| Family | Best for | Latency | Cost |
|---|---|---|---|
| Claude Opus 4.7 | Deep multi-step reasoning, War Room synthesis, long-context analysis. | 2–6s | $$$ |
| Claude Sonnet 4.6 | Most agent work. Great default. | 1–3s | $$ |
| Claude Haiku 4.5 | Routing, simple classifications, voice agents (sub-second). | 0.4–1s | $ |
Pick per agent
Each agent / team picks its own model. Mix and match: route the cheap stuff to Haiku, hand the hard stuff to Opus.
Prompt caching
System prompts (agent identity + skills + tool catalog) are cached server-side. Repeated calls inside a 5-minute window hit the cache and pay 1/10 the input-token price.