Agent Cost Strategy

Agent Cost Strategy

Use the cheapest model that can reliably do the job. Most tasks don't need your most powerful model.

The Three Tiers

Task → Tier Routing

Fix failing tests          → Fast/Cheap
Write boilerplate          → Fast/Cheap
Research / search          → Fast/Cheap
Cron / scheduled tasks     → Fast/Cheap (always)
Short replies (hi, ok)     → Fast/Cheap (always)
Background monitoring      → Fast/Cheap (always)
Build new feature          → Mid-tier
Review a PR                → Mid-tier
Main assistant dialogue    → Mid-tier (default)
Architecture decisions     → Powerful
Deep code review           → Powerful
Stuck after 2 attempts     → Escalate one tier up

Heartbeat / Cron Model Rule

Always specify the cheapest model for scheduled and background tasks — they run frequently and costs add up fast. Check your platform's config for how to set a model per cron/heartbeat job.

For heartbeat intervals: set them just under your provider's cache TTL to keep the prompt cache warm and pay cache-read rates instead of full input rates. Check your provider's docs for the exact TTL.

Communication Pattern Rule

One-word and short conversational messages (hi, thanks, ok, sure, yes, no) should always route to Fast/Cheap. Never burn a mid-tier or powerful model on an acknowledgment.

Cache Optimization

Prompt caching cuts costs 50-90% on repeated context. Cache writes cost ~25% more but pay off after just 1-2 reuses. See references/cache-optimization.md for patterns and break-even math.

Batch API (Non-Urgent Tasks)

For cron jobs, scheduled analysis, or anything that doesn't need an immediate response — use the Batch API (Anthropic/OpenAI both offer it). 50% discount in exchange for async delivery (results within 24h). Never use real-time API for background work that can wait.

Sub-Agent Model Rule (Critical)

Always explicitly set the model when spawning sub-agents. Never rely on defaults — the default inherits the parent session model (expensive mid-tier). One month of sub-agents defaulting to Sonnet = 96% of costs going to Sonnet when it should be split ~80/20 Haiku/Sonnet.

sessions_spawn → always include model: "claude-haiku-4-5-20251001" (or equivalent fast-cheap)

Every sub-agent task that can be done by Haiku must use Haiku. No exceptions.

New Session / Machine Cold Start Cost

When starting a fresh session (new machine, new session after /new), the cache is empty. The first few messages will write the entire context (skills, workspace files, memory) to cache at 1.25x the normal input rate. This is unavoidable but temporary — it pays off within 2-3 messages once the cache warms up.

Don't panic at the first few messages being expensive on a new machine. The cache write cost is a one-time investment that makes every subsequent message ~90% cheaper.

Signs You're Over-Spending

• Running powerful models on tasks Fast/Cheap can handle
• No caching on repeated system prompts
• Heartbeat/cron jobs using the default (expensive) model
• Sub-agents spawned without explicit model = biggest cost leak

Session & Cache Management

Never delete or end sessions to "save money" — it backfires.

Anthropic's prompt cache builds from repeated context within a live session. When a session starts fresh, all context (system prompt, workspace files, skills) loads cold — typically 400-600k tokens at full cost. Once cached, subsequent messages cost ~10% of that.

The math:

• Cold session start: 600k tokens × full price = expensive
• After cache warms up: 600k tokens × 10% cache price = ~90% cheaper per message
• Deleting a session destroys the cache and forces a full cold reload next time

Rules:

• Let sessions run as long as possible
• Only start a new session (/new) when context is genuinely full (>80%)
• Never tell a user to delete conversations — it spikes costs
• The longer a session runs, the cheaper each message gets

Delegation rule (keep main agent lean):

• Main agent (Sonnet/mid-tier) = conversational only: planning, coordination, reviewing results
• Sub-agents (Haiku/fast-cheap) = all actual doing: file edits, research, builds, data tasks
• Keeping the main agent conversational reduces its context growth and keeps cache hits high