Author(s): Udaykiran Estari

Originally published on Towards AI.

The Wrong Question Is ‘Which Claude Model Should I Use?’

The fastest way to waste money in Claude Code is not picking the expensive model. It is leaving effort unconstrained, watching a normal coding session turn into a 7x token multiplier or a 1,000-subagent workflow. Model choice sets the capability ceiling, but your effort configuration dictates how aggressively you burn toward it. To stop bleeding budget, we need a practical, hands-on routing policy that treats model tier and reasoning depth as separate, governable levers.

The article argues that controlling Claude Code costs is less about choosing the “best” model and more about treating model tier and effort level as separate levers. It explains that the common failure modes come from pairing a big model with too little effort (starving it), assuming smaller models can’t do serious work (when sufficient effort matters), or using the most expensive cell when both reasoning capacity and deep exploration are needed. It then details how effort drives token volume and agentic orchestration—often causing large multipliers and runaway budgets—while noting edge cases like Haiku 4.5 not supporting the newer effort parameter and the importance of prompt caching, which can make a typically “expensive” model cheaper for repeated workloads. Finally, it proposes a practical governance approach: start with a reasonable default (e.g., Sonnet 5), cap effort by role/workflow, escalate only when diagnostics show context is correct but work is too shallow or reasoning fails, and reserve ultracode as budgeted infrastructure rather than a convenience toggle.

Read the full blog for free on Medium.

Join thousands of data leaders on the AI newsletter. Join over 80,000 subscribers and keep up to date with the latest developments in AI. From research to projects and ideas. If you are building an AI startup, an AI-related product, or a service, we invite you to consider becoming a sponsor.

Published via Towards AI