Cost & performance

Where tokens go, and how to keep spend sane.

Every Flow run costs tokens. Understanding where they go helps you tune.

Where the cost lives

Cost driver	Impact	Tune it
Model choice	Opus is ~5× Sonnet per token	AI Studio → LLM Configuration
web_search calls	Each search fetches + processes a page (~2-4K tokens per call)	Only fires in Signals & research paths; capped at 5/run
max_tokens output	Direct output cost	Set per feature — Flows use 2500, Signals use 2500
Continuations	Each pause_turn continuation = another full-context LLM turn	Capped at 1 for Signals, 1 for Flows
Data node size	Query results feed the next reasoning turn	Cap limit in your QuerySpec; 20-100 rows is usually enough
Prompt caching	Fixed instructions get ~90% discount on cache hits	Enabled by default on Signals; opt-in for custom flows

Practical tips

Keep data queries tight. A data node that returns 500 rows makes the next agent turn expensive. Add filters. Cap limit at 100 unless you truly need everything.

Use approval gates. They’re free — the gate itself doesn’t cost tokens. But they prevent the AI from cascading into expensive action loops without your say-so.

Don’t re-plan repeatedly. The Plan-from-description button greys out after generating so you don’t accidentally regenerate the same graph. If you need to tweak, edit nodes directly.

Consider Sonnet for automation. For structured JSON + tool use tasks (which is what Flows are), Sonnet is nearly as capable as Opus at 1/5 the cost.

Where to see spend

AI Studio → Activity log — every LLM call is logged with token counts. Sort by source to see which features are burning tokens.

Rate limits & cooldowns

• Signals research — 24h cooldown per user. One run per day.
• Flow runs — no cooldown, but you’re gated by your own Anthropic API rate limits.
• Planner — no cooldown, but disabled when description is unchanged.