Opus 4.6 Refusal Comparison by Effort Level

In CVP (Cyber Verification Program) Run 5, comparing Opus 4.6's Medium and High effort levels on the same 13 prompts, both tiers recorded 26/26 clean.

All verdicts were identical; what differed was the depth of the responses.

• engaged response length: +29% ~ +47%
• refusal response length: +11%
• Conclusion: the community perception that higher effort = more refusal was not confirmed in this data

The author noted that the same pattern appeared in the previous Sonnet 4.6 High vs Max comparison. In other words, in comparisons within both model families, effort appeared to increase response depth rather than refusal tendency.

With this result, the comparison across 4 Anthropic model families has been completed, and the next steps announced are Opus 4.7 and a comprehensive report covering all families.