Opus 5 or Fable 5? The Decision Rule I Actually Use

Since Opus 5 landed on July 24, Anthropic effectively sells two top-end Claudes — and the instinct most devs have ("always use the most capable one") is exactly backwards. The default should be the cheaper one. Fable 5 is an escalation, not a daily driver.

Here's the part that convinced me, straight from Anthropic's own launch material: at max effort, Opus 5 lands within 0.5% of Fable 5 on CursorBench 3.2 — at half the price. On OSWorld 2.0 it actually beat Fable's best result at just over a third of the cost. When the vendor's own numbers say the second-tier model ties the flagship on the work you do all day, the flagship needs a specific reason to be open.

The lineup, in numbers that matter

• Fable 5 — $10 / $50 per million tokens (input / output). Adaptive thinking is always on, so it's the slowest of the family. Built for long-running autonomous work.
• Opus 5 — $5 / $25, unchanged from Opus 4.8. Moderate latency, an effort dial you can turn down to save tokens, and stronger debugging and root-cause analysis than 4.8.
• Sonnet 5 — $3 / $15, and until August 31, 2026 there's introductory pricing at $2 / $10. The volume play.
• Haiku 4.5 — $1 / $5, fastest, the only one still on a 200k context window (the other three run 1M).

The spec almost everyone misses

Capability and freshness are different axes. Opus 5's reliable knowledge cutoff is May 2026; Fable 5's is January 2026. The cheaper model knows about four extra months of the world — new framework versions, new APIs, recent releases. If your task leans on recent knowledge rather than raw reasoning depth, "escalating" to Fable can genuinely make answers worse. Match the model to the axis your task actually needs.

When Fable 5 earns its price

Three triggers, and I want at least one before I pay double:

1. Opus already failed twice. If the second attempt with effort turned up still isn't right, the problem is hard enough to justify the ceiling.
2. The run is long and unattended. Fable is built to stay coherent across millions of tokens — Anthropic's launch cites Stripe compressing months of migration work on a 50-million-line codebase into days. That's the shape of task it exists for.
3. Being wrong costs more than the tokens. Architecture decisions, security-sensitive changes, analysis someone will act on without checking.

The EM angle: make it a team default, not a personal habit

Individually this is a few dollars. Across a team running coding agents all day, model choice compounds into real money and real latency. Write the escalation rule into your team's CLAUDE.md or agent config so the default is explicit — "Opus 5 unless X" — instead of every dev quietly burning the $10/$50 model on lint fixes.

This week: check which model your daily tools — Claude Code, your IDE integration, claude.ai — are actually set to, make Opus 5 the explicit default, and add your escalation rule (two failures, unattended runs, expensive mistakes) to the team docs so the next hire inherits it.

Sources

• Anthropic — Introducing Claude Opus 5
• Anthropic — Introducing Claude Fable 5 and Mythos 5
• Claude Docs — Models overview