Claude Opus 5’s Verbosity Outruns Standard Benchmarks

The em-dash count is a funny but revealing proxy for a real developer complaint: capability gains can arrive alongside worse output discipline.

• –Punctuation frequency measures style, not intelligence, and varies with prompts, system instructions, and task mix.
• –Anthropic confirms Opus 5 tends to produce longer visible responses than prior Opus models.
• –Developers should add verbosity, formatting, and instruction-adherence checks to model regression suites.
• –Explicit concision prompts help, but Anthropic says effort controls reasoning more reliably than final response length.
• –Better evaluations pair task success with token use, latency, verbosity, and user preference.