Skills can partially substitute for model scale
Claude model family performance. A weaker model with skills consistently outperforms the next-tier model without skills.
Without Skills
With Skills
Substitution
0
10
20
30
40
50
Haiku+skills beats Sonnet
Sonnet+skills beats Opus 4.5
11.0
17.3
22.0
30.6
27.7
31.8
45.3
44.5
Haiku 4.5
Sonnet 4.5
Opus 4.5
Opus 4.6
Model (increasing capability →)