Skip to content

When does routing a bounded task to a cheaper model via API pay off?

Last updated 2026-08-17 · Confidence: documented — Anthropic’s and Google’s pricing pages and Claude Code’s authentication policy; the $20–25 single-run figure is a call participant’s first-hand report (tried).

Cheap API tiers price tokens at roughly 1/10th to 1/30th of frontier rates, so a heavy run that burns $25 on a frontier model costs about a dollar once the task is bounded enough for a cheaper model — and a personal internal tool may not need API billing at all.

Current per-million-token prices, from the Anthropic and Gemini pricing pages (input/output):

  • Claude Fable 5: $10 / $50
  • Claude Sonnet 5: $2 / $10
  • Claude Haiku 4.5: $1 / $5
  • Gemini 3.7 Flash: $0.75 / $3.75 (promotional through Dec 2026, then $1.50 / $7.50)
  • Gemini 3.5 Flash-Lite: $0.30 / $2.50

A single analysis run reported at $20–25 in frontier tokens (tried — first-hand, 2026-08-14) re-priced at the Flash-Lite or Haiku tier lands near $1; the Batch API halves Claude prices again for non-urgent jobs. Against very low labor costs — staff at ~$3/hour came up on a call — a $25 frontier run rivals a day’s wages, while the routed version costs minutes of them.

The gap only cashes out on bounded tasks: a fixed, well-specified job (extract, classify, screen, reformat) whose output is easy to verify, repeated enough times that per-run cost matters. Use the frontier model to design and spot-check the pipeline; route the volume through the cheap tier.

Pro/Max usage limits “assume ordinary, individual usage of Claude Code and the Agent SDK” (legal and compliance) — a scripted claude -p run uses your normal subscription login (its --bare mode is the exception: it reads no OAuth credentials and needs an API key). So a tool you run yourself can ride flat-rate subscription tokens. The same page draws the line: developers building products or services for others, “including those using the Agent SDK”, should use API keys — routing other people’s requests through Free/Pro/Max credentials is not permitted. A screening tool your whole team hits belongs on the API, which is exactly where cheap-model routing matters.