Verdict
Built from the measurements below
- Accuracy is too close to call: Claude Haiku 5.5 scored 68.8% and Claude Sonnet 5.5 78.8% on the same 80 questions, a gap within what chance alone could produce.
- Claude Haiku 5.5 is 11× cheaper per question as billed ($0.00048 vs $0.0054), reasoning tokens included.
- Claude Haiku 5.5 generates text 1.5× faster: 192 tokens per second against 126.
Which should you choose?
Choose Claude Haiku 5.5 for the lower cost per correct answer ($0.0007 against $0.0069) and faster streaming. Choose Claude Sonnet 5.5 for shorter outputs (about 456 tokens per question against 873). The widest subject gap is in engineering: Claude Sonnet 5.5 answered 13 of 15 correctly, Claude Haiku 5.5 10 of 15. With so few questions per subject, treat this as a lead to test on your own prompts, not a result.
Accuracy gap -10.0 points (z = -1.44). We call a winner only when z is at least 1.96, the usual 95% threshold. Speed, latency and price need a 10% margin.