Skip to content- Sources: primary, discussion
- Summary: On 1,000 synthetic identities from pseudopeople, Claude Opus 5.5 scored 99.0% at $3.86 per 1,000, while Jev alone scored 95.7% at $0.05. Sending Jev's 30 to 90 percent confidence cases to Opus scored 98.8% at $0.88, and on a harder set it scored 95.3% versus 95.1% for Opus at $2.94 versus $5.39. GPT-6 Luna was more accurate alone but overconfident, so its hybrid reached only 91.6% on the harder set.
- Why it matters: Confidence calibration, not raw accuracy, decides whether a cheap first-pass model works in an escalating LLM judge pipeline.