6 setups, 28 tasks: Haiku alone cuts Claude costs 93%, delegating to cheaper models costs more

scomage · reddit · 2026-10-22

To stretch his Claude quota, the author built an MCP tool to hand Claude Code's mechanical coding work to cheaper models, then measured 6 setups across 28 real tasks. Most expectations were wrong:

Key results (median per run)

Haiku 5.5 alone: passed all 84 checks at $0.008—93% cheaper than Sonnet 5.5 alone ($0.113). Haiku used more tokens (297k vs 152k, mostly cache reads) but its tokens are far cheaper.
Handing off to outside models cost more: Sonnet + DeepSeek V4.1 Flash or a local 27B via MCP ran $0.113–0.156—the spec-writing, waiting, and review turns dominate; the worker's bill (1¢) can't offset that.
Subagents are the cheap handoff: shared session, no spec/review round trip, billed at Haiku rates; savings grow with task size.
Opus 5.5 as orchestrator cost about the same as Sonnet alone.

Quality blind-grade (10 real backlog items, Opus-graded vs merged commits): Opus alone scored 99 and was the only setup never needing major fixes at 20x Haiku's cost; Haiku alone 94 (mostly minor fixes); Sonnet+Haiku subagent 90.

Limits: ready-made specs (spec-writing not counted), Python only, Opus as grader may favor its own code, and read-heavy offloading untested—where other tools report large savings. Tasks, raw data, and repro steps are open-sourced at github.com/gabrielqfr/skrode.

Original post →

6 setups, 28 tasks: Haiku alone cuts Claude costs 93%, delegating to cheaper models costs more

More from coding & agent