Claude's newest model is a step forward and two steps back, and it's infuriating
This title could be clearer and more informative.Try out Clickbait Shieldfor free (5 uses left this month).
Claude Opus 4.7, Anthropic's latest flagship model, brings genuine improvements in instruction-following — it no longer rewrites code you didn't ask it to touch — but these gains come with notable regressions. The model hallucinates more than its predecessor, avoids proactive web searches even when clearly needed, gives shallower responses by default, and burns through up to 35% more tokens due to an updated tokenizer. Real-world testing and Reddit reports corroborate that the model feels lazier and requires more hand-holding, making the upgrade feel like a mixed bag for everyday users.
Table of contents
Opus 4.7 is Anthropic's most capable model yetOpus 4.7 finally does what you tell it toUnfortunately, the better listening comes at a costOpus 4.7 is also somehow... lazier?Sort: