Price per Token Is Not Cost per Task: The Claude Developer Blog's Model-Selection Playbook

A higher price per token does not always mean a higher cost per task. That is the working rule in Addy Osmani’s model-selection playbook on the Claude developer blog, which covers the four generally available Claude models: Haiku 5.5, Sonnet 5.5, Opus 5.5 and Fable 5.2.

List prices per million input and output tokens run $0.10 and $0.50 for Haiku (on prompts up to 100,000 tokens), $2 and $10 for Sonnet, $4 and $20 for Opus, and $8 and $40 for Fable. Osmani’s default is to start most work on Opus 5.5 at medium effort — Claude Code’s default. Haiku suits high-volume classification, routing, extraction and simple subagents. Sonnet suits well-scoped coding and chat. Opus handles complex coding, computer use and long agentic loops. Fable is for major refactors, architecture, deep research and the hardest problems.

The benchmark results are where the price list stops being a guide. On Terminal-Bench 3.0, Opus 5.5 at high effort scored 58.9% using half the tokens of Fable 5.1 at max, which scored 58.0%. Fable 5.2 at high scored 45.3% on FrontierCode 2.0, nearly matching Opus 5.5 at max at roughly half the cost per task. And on SWE-bench Pro, running tasks at low effort and re-running only the failures at high passed about 97% for $0.17 each, versus 95.3% for $0.29 — a higher pass rate at about 41% less per task.

Anthropic’s own Fable 5.2 release says the same from the vendor side: it still recommends Opus 5.5 as the daily driver, with Fable as an advisor or orchestrator for smaller models. The saving shows up in production, too. SaaStr’s Jason Lemkin reports that moving SaaStr Connect to Haiku 5.5 and a cheaper model mix cut total costs 72% for the same workflow.

Model routing is an engineering decision with a measurable outcome. Track pass rate and dollars per completed task on your own workloads — the playbook’s own advice — before you ship an agent on the most expensive model by default.