Chen Deng

~/log / match-the-model-to-the-task

Match the model to the task: how my SEO agent went from $560 to $11 a year

#ai-ops #cost #seo

My SEO agent rewrites metadata, generates alt text, injects structured data, and writes three blog posts a week for two production sites. It originally ran on Claude, and the AI bill came to roughly $560/year.

Then I asked an uncomfortable question: how much of this work actually needs a frontier model?

The honest answer was almost none of it. The agent’s tasks are production work — high-volume, well-specified, checkable:

  • Rewrite this meta description under 155 characters, targeting this keyword
  • Generate alt text for this product photo
  • Draft a blog post from this keyword brief and outline

None of that requires deep reasoning. It requires following a spec, and cheap models follow specs fine. I swapped the brain to DeepSeek and the annual AI cost went from ~$560 to ~$11. I watched rankings and click-through for regressions. There weren’t any.

The general principle

Agent work splits into two kinds:

  • Judgment work — deciding strategy, evaluating trade-offs, anything where a wrong answer is expensive and hard to detect. Use the best model you can get.
  • Production work — executing a defined task at volume, where output quality is easy to verify. Use the cheapest model that passes your quality bar.

Most agent systems I see get this backwards: they burn frontier-model tokens on production work because that’s the model the developer likes, and the cost makes the whole system feel unviable. Across my fleet the pattern holds — Claude does strategy and proposals, commodity models do the volume, and the marginal cost of an agent-run business stays near zero.

The uncomfortable question is worth asking of every agent you run: is this judgment, or is this production?

Thoughts, disagreements, or building something similar?Email me — the whole point of this log is finding people who care about this stuff. Subscribe viaRSS.