Models
The Claude models a company's agents can think with. People pick one per agent; every reply and job runs on it and is billed at its price.
| Model | API id | Input | Output | Fallback |
|---|---|---|---|---|
Claude Opus 5.5DefaultThe most capable. Best for building apps, long multi-step work and hard decisions. | claude-opus-5-5 | $4 /M | $20 /M | Server-side |
Claude Sonnet 5.5Fast and capable at half the cost. Good for everyday work and most chats. | claude-sonnet-5-5 | $2 /M | $10 /M | Server-side |
Claude Haiku 5.5The quickest and cheapest. For simple, high-volume replies. | claude-haiku-5-5 | $0.10 /M | $0.50 /M | None |
Prices per million tokens. Cached reads cost less; Syvna meters what each call actually used.
Agent harness
Model per agent
Each agent picks Opus, Sonnet or Haiku on its page. Agents without one follow CLAUDE_MODEL.
Effort
Per agent, low to max, on its page. Always stated on the call, because the models' defaults differ; Syvna's default is medium.
Reasoning
Agents think first by default. Sonnet and Haiku agents can answer directly between steps, at high effort or below; Opus always thinks.
Refusal fallback
A safety-classifier decline is re-served inside the same call on Opus and Sonnet. Haiku has none.
Prompt caching
Rules and tools cached an hour across companies, the agent's identity an hour, the history per call.
UI bar
Agents that build apps read the template's design guide before any page, and check desktop and mobile before they reply.