Which Claude model runs each AI feature

Used for: Persona generation, Stage suggestions, Action plan generation. These are repetitive, structured tasks where Haiku produces the same quality as Sonnet for ~3× lower cost.

Pricing: $1 per million input tokens, $5 per million output. Cached reads cost ~10% of regular input.

Used for: Insights synthesis (combining all stages, personas, and pain points into a single strategic narrative). Sonnet is meaningfully better than Haiku for this kind of cross-context reasoning.

Pricing: $3 per million input tokens, $15 per million output.

By default, document import uses ANTHROPIC_MODEL from the environment (defaults to Opus 4.7 for backwards compatibility with the original implementation). For cost-sensitive deployments, override with Haiku — quality is acceptable for most CX docs.

All AI routes use Anthropic's prompt caching: the system prompt and stable journey context are cached for ~5 minutes after each call. Re-running the same feature on the same journey within that window saves ~70% of input cost and ~50% of latency.

Anthropic's free tier is 5 RPM per model. If you hit it, Journey Builder waits 15 seconds and retries automatically — you'll see a brief "Rate limited" toast but the call usually succeeds. Heavy users on a paid Anthropic plan rarely see this.