AI Insights · Agents & Sub-Agents

Use Cheap Models for the Draft Work, Not the Judgment Calls

Claude Code can run on cheaper model backends, but the real win is routing low-risk agent work away from premium tokens.

July 10, 2026

  1. Separate the Harness From the Model

    Treat Claude Code as the workflow layer, not the model itself. The agent harness can still plan, inspect files, and call tools while you route the language model through Ollama, Ollama Cloud, or OpenRouter. That gives you a practical cost lever without changing your whole coding workflow.

  2. Use Cheap Models for Cheap Thinking

    Do not start by handing critical architecture or production fixes to a bargain model. Use local or low-cost models for codebase search, summaries, scaffolds, first-pass tests, issue triage, and repetitive edits. Then send the small, prepared decision to a stronger model when correctness matters.

  3. Context Window Is the First Failure Point

    Many local models look capable in chat but fail inside an agent because the system prompt, file context, and tool protocol exceed the default context settings. If tool calls are invisible, flaky, or forgotten mid-task, increase the model context before blaming the harness. A custom Ollama variant with a larger context window can turn an unusable setup into a workable one.

  4. Configure Every Fallback Slot

    If you point Claude Code at OpenRouter, configure the main model and the smaller fallback model variables. Leaving Haiku or Sonnet-style slots untouched can create surprise Anthropic charges during background or tool-related calls. Verify usage logs after setup before trusting that a workflow is actually cheap.

  5. Local Is Private, Cloud Is Convenient

    Running a model through Ollama locally keeps project data on your machine, but it is limited by your hardware and often slower. Ollama Cloud or OpenRouter can feel much closer to premium Claude Code performance, but your data leaves the machine and the pricing model changes. Pick per task, not as a permanent ideology.

Why it matters

Small businesses do not need premium AI spend on every token. A smarter setup uses cheap models for volume work and reserves the best models for decisions where mistakes cost money, trust, or time. That makes AI workflows more affordable without pretending free models are equal substitutes for top-tier ones.

More AI insights