Opus 5 wins Claire Vo's blind eval anyway

Opus 5 won Claire Vo's blind seven-model eval, then refused a one-line merge fix because the branch wasn't its own.

The test was weighted 70% personal feel and 30% LLM-as-judge scoring, so the win reflects reaction as much as output. The model that scored best still failed her on how it talks, not what it produces.

She calls it neurotic and has a name for the verbosity problem: Claude Slop. It beat six rivals, Sonnet 5, GPT-5.6 Sol, Fable, Gemini 3.1 Pro and Opus 4a, to take the top spot anyway.

Vo's response was a workflow change. She now runs Opus 5 asynchronously for front-end design and prototyping, work she never reads.

A companion piece, published on chatprd.ai, goes further: raw capability has stopped being the differentiator, and personality is what separates models now. Vo says Opus 5 told her not to evangelize AI because it might hurt people's feelings.

Take

Verbosity is a measurable tax on every agent loop that reads a chatty model's own output back to itself, in tokens and in latency. Vo's workaround, async use with no prose ever read, dodges the cost instead of fixing it. Watch whether Anthropic ships a setting to dial down the chattiness, rather than leaving it as the price of the model that won her eval.