Verdict
Built from the measurements below
- Accuracy is too close to call: Claude Sonnet 5.5 scored 78.8% and Gemini 3.8 Flash 82.5% on the same 80 questions, a gap within what chance alone could produce.
- Gemini 3.8 Flash is 1.5× cheaper per question as billed ($0.0037 vs $0.0054), reasoning tokens included.
- Gemini 3.8 Flash generates text 1.1× faster: 142 tokens per second against 126.
- Claude Sonnet 5.5 starts answering sooner: 0.7 s to the first token, against 3.2 s for Gemini 3.8 Flash.
Which should you choose?
Choose Claude Sonnet 5.5 for a quicker first token, shorter outputs (about 456 tokens per question against 925) and more hosts to choose from on OpenRouter (5 against 2). Choose Gemini 3.8 Flash for the lower cost per correct answer ($0.0045 against $0.0069) and faster streaming. The widest subject gap is in law: Gemini 3.8 Flash answered 6 of 6 correctly, Claude Sonnet 5.5 3 of 6. With so few questions per subject, treat this as a lead to test on your own prompts, not a result.
Accuracy gap -3.7 points (z = -0.60). We call a winner only when z is at least 1.96, the usual 95% threshold. Speed, latency and price need a 10% margin.