60x. That's how much one developer cut their Claude bill by redirecting routine tasks to a smaller model - without touching any of the work that actually needed Sonnet.
The realization came from auditing usage logs: the biggest cost driver wasn't the hard stuff. It was file classification, JSON reformatting, pulling fields out of text, summarizing documents they were going to skim anyway. All of it billed at the same rate as genuinely complex reasoning work.
The obvious fixes didn't hold up. Switching to Haiku (Claude's faster, cheaper tier) helped, but at high volume the savings were still marginal. Tighter prompts moved the needle a little. The /compact command - which compresses conversation history to reduce how many tokens the model needs to process - delayed the problem rather than solving it. None of these approaches fixed the core mismatch: a frontier model handling tasks that a much cheaper model could do equally well.
The actual fix was routing. A small side model handles all bulk, repetitive processing. Sonnet only runs when the task requires real judgment. The result was that 60x cost reduction.
The Tasks You're Probably Overpaying For
Most people using Claude for real work fall into the same pattern. The expensive tasks are memorable. The cheap ones happen hundreds of times quietly in the background.
Things that almost never require a frontier model:
- Classification - sorting items into categories, flagging records for review
- Extraction - pulling named fields from structured or semi-structured text
- Reformatting - converting between JSON, CSV, plain text, markdown
- Simple routing decisions - yes/no calls like "does this need a human?"
- Light summarization - collapsing a 200-word paragraph to 40 words
The useful test: can you describe the rule explicitly? If yes, a smaller model will follow it just as reliably at a fraction of the cost. These tasks require pattern matching, not reasoning.
Putting It Into Practice
For developers hitting the Claude API directly, this means adding a routing layer before your calls. A small model checks task type; simple work goes to a cheaper endpoint, complex work proceeds normally. For Claude Code users, the practical version is being deliberate about what you ask the agent to do - batch your mechanical tasks and run them separately from the reasoning-heavy ones.
The price differences between model tiers are significant at volume. Claude Haiku processes tokens at a fraction of Sonnet's cost, and the pattern holds across providers - GPT-4o mini versus GPT-4o shows similar gaps.
Anyone running automated workflows that touch Claude regularly should pull their usage logs first. The breakdown between "needed Sonnet" and "didn't need Sonnet" is almost always more lopsided than expected.