Opus 5 wins 8 of 13 rows on Anthropic's own launch chart, at half Fable 5's price. That flipped how I route work across my agents. This page has the exact rule, plus the one-page cheat sheet.
→Opus 5 dropped July 24 at $5 in, $25 out per million tokens. Same price as Opus 4.8, near Fable 5 results on agentic work, and it's the new default on Claude Max.
→One benchmark row explains the split. Humanity's Last Exam with no tools: Fable wins. Same exam with tools: Opus 5 wins. Fable has the deeper raw brain. Opus 5 is better at using its hands.
→The safety story flipped too. Opus 5 scored most aligned in Anthropic's behavioral audit, but at developing exploits the Mythos tier is still 3x ahead, 13 to 4. The frontier stays gated for a reason.
🔒Drop your email and I'll send the routing cheat sheet
The one-page Claude routing rule: which model gets which job, what the effort dial replaces, and when a Fable 5 session is worth 2x the price. Plus whatever I learn running 20+ agents each week.
Join 3,000+ builders already getting this. No spam.
THE ROUTING RULE — 4 MOVES
01
Opus 5 becomes the daily driver
Code, subagents, long agent runs, computer use. It ships at the same price Opus 4.8 did and it's included flat on a Max plan. Within half a percent of Fable 5 on agentic coding benchmarks. Anthropic's own launch table below. Default everything here first.
Opus 5 $5 / MTok in $25 / MTok out
Fable 5 $10 / MTok in $50 / MTok out
Sonnet 5 $2 / MTok in $10 / MTok out (intro pricing)
02
Crank the effort dial before switching models
Opus 5 has an effort setting: low, medium, high, xhigh, max. Medium handles most work. A hard bug gets high or max before it gets a more expensive model. That dial kills most of the reasons people used to reach for the flagship. Anthropic's chart below: score per dollar by effort level, and the Opus 5 curve sits above everyone.
03
Fable 5 is the architect
It keeps the deeper raw brain: system design, big planning calls, orchestration decisions, taste. The tell is one benchmark row. Humanity's Last Exam with no tools: Fable wins. With tools: Opus 5 wins. But Fable bills usage credits at API rates even on Max, so it runs in short fresh-context sessions and has to earn every call. Escalate only after max-effort Opus 5 fails.
04
Grab the one-pager
The whole rule on one page: three models, their jobs, their prices, the escalation ladder. Save it, and send it to whoever still routes everything to the most expensive model out of habit.