Claude web client behaves differently than API on same model

Same Claude model, two surfaces, different behavior. Side‑by‑side tests with thinking disabled on both still show the web client giving odd first‑person/plural phrasing and more constrained answers, while the API returns normal completions. That shifts the blame toward the web layer or prompts, not the base model.

The fallout turned into workflow pain: refusals blocking routine coding, slower agent loops, and extra cost. One report burned $100+ in an hour, another clocked a simple posting task at ~70 seconds via Claude vs ~5 seconds manually. Anthropic says some Opus 4.7 malware‑style refusals came from a stale safety prompt in older builds and that fixes are being worked on.