Claude Opus 5 Review

Opus 5 is the strange one: newer than 4.8, priced identically, and harder to recommend. The headline change is that thinking is now on by default, which sounds like an improvement and is, until you meet the failure modes on the other side of the switch. Explicitly disable it and the model will occasionally write a tool call into its visible text instead of emitting a real call, so the turn succeeds, nothing runs, and no error is raised anywhere.

In an agentic loop that silent failure poisons every later turn. It will also leak internal reasoning tags into responses. Disabling thinking is only accepted at high effort or below, so even the escape hatch is gated. Operationally it sits worse than its predecessor too: excluded from priority capacity entirely, and fast mode runs at double the standard rate. The 1M window, the $5 and $25 pricing and the raw reasoning are all genuinely good. But a model whose own guidance is to avoid the setting it just made configurable has not earned better than a B.