Anthropic CEO drops 12-page PDF on using Sonnet 5.5 vs Opus 5.5
msharmas · x · 2026-10-04
Anthropic CEO Dario Amodei released a 12-page PDF on pairing Sonnet 5.5 and Opus 5.5, praised as "more useful than most paid AI courses." It lays out a 12-step blueprint for a faster, cheaper, more capable Claude workflow:
- Pick models by task, not prestige: judge how clearly the task is defined and how costly a wrong answer is;
- Write the finish line first: deliverable, constraints, evidence, allowed actions, and the exact check proving completion;
- Package one clean brief: goal + current state + sources + acceptance test, instead of dumping the whole chat;
- Division of labor: well-scoped execution to Sonnet 5.5; ambiguous architecture, sprawling migrations, high-stakes judgment to Opus 5.5;
- Sweep effort before switching models: Sonnet-high and Opus-medium are different defaults.
A companion thread adds the cost reality: Sonnet is cheaper per run, but Opus can be cheaper per finished task — the difference lies in effort levels, cache misses, and retries you paid for but never shipped.
More from coding & agent
- RLM explained: MIT's recursive REPL approach lets LLMs handle contexts beyond the window — AnuranBuilds · 2026-10-04
- Open-source workflow extends static video shots with no degradation using masked bridge resampling — roychodraws · 2026-10-04
- After 18 years of hand-writing code, this dev quit for good — PRs up 50% with no more bugs — bibryam · 2026-10-04
- Perplexity's Decisions API beats Pokémon FireRed's Elite Four in one shot for $0.028 — AravSrinivas · 2026-10-04
- Hermes adds /blueprint: 16 built-in templates for recurring agent automations — Teknium · 2026-10-04
- Delvetown day four: AI agent town produces poems, games and SBF fanfiction — max_paperclips · 2026-10-04