Effort levels

A per-request dial on Claude models controlling how much reasoning/compute is spent: low → medium → X-high → max. Higher effort = better output on hard tasks but more tokens, more cost, and more latency. Choosing the level is routing within a single model.

Guidance from the Fable 5 launch

• Start low. Both creators converge: for Fable 5, begin at the lowest effort — it's usually sufficient, and higher levels are often overkill (Berman found even his lowest setting felt "too high").
• Rough equivalence: Nate Herk pegs Fable on low ≈ Opus 4.8 on X-high (and cheaper).
• Diminishing returns at the top: Nate sees little difference between X-high and max on Opus 4.8; Anthropic bumped max to be a small improvement over X-high on Fable.
• Cognition's Frontier Code eval: Fable 5 scores highest among frontier models even at medium effort.

Sources

• MYTHOS MYTHOS MYTHOS (Matthew Berman)
• Claude Mythos is Finally Here (Nate Herk)