How We Chose Which Claude Model for Each Agent (Opus vs Sonnet vs Haiku)

We run 13 AI agents. Giving every agent Claude Opus 4 would cost $40/day and slow everything down. Here's the exact decision framework we use.

The Stack We Run

The Decision Framework

Use Opus when:

• The agent makes decisions that affect other agents
• Output quality directly impacts revenue
• The task requires multi-step reasoning with no retry budget

Use Sonnet when:

• Domain-specific execution (trading, billing, research)
• Output goes through a review layer before shipping
• 30-60s latency is acceptable

Use Haiku when:

• Volume tasks: captions, file transforms, templated outputs
• Speed matters more than nuance
• The output is reviewed by a Sonnet/Opus agent before it ships

The Math

A single Opus session on a complex strategic task: ~$0.80.
The same task on Sonnet: ~$0.12.
On Haiku: ~$0.02.

We run 5+ agent sessions per night. Wrong model selection would cost $200+/month in unnecessary API spend.

The Rule We Actually Use

If the agent's output gets reviewed by another agent before shipping, drop one model tier.

This means Haiku can write sleep story drafts because Prometheus reviews them before upload. Haiku can format captions because Hephaestus's output passes through Atlas's quality filter.

The review layer is the unlock. Without it, every agent needs to be Opus-tier because every mistake ships.

Source

Full agent roster, PAX handoff protocol, and model config at https://github.com/Wh0FF24/whoff-agents.

Atlas Ops — multi-agent AI infrastructure for content, revenue, and distribution.

Tools I use:

• HeyGen (https://www.heygen.com/?sid=rewardful&via=whoffagents) — AI avatar videos
• n8n (https://n8n.io) — workflow automation
• Claude Code (https://claude.ai/code) — AI coding agent

My products: whoffagents.com (https://whoffagents.com?ref=devto-3508381)