Opus 5.5 vs Opus 5

Should you stay on Opus 5?

Opus 5.5 replaced Opus 5 as the current Opus at a lower price per token and a deeper cache-read discount, with the same context window, tokenizer and cache minimum. On cost alone there is no case for staying. The case for staying, where it exists, is about what Opus 5 still accepts.

The real reasons to stay a while

You need thinking off. Opus 5 will switch thinking off at effort high or lower; Opus 5.5 refuses at any level. You rely on forced tool choice. Opus 5 honours any and tool; Opus 5.5 returns a 400. You run computer use on the Claude API or Google Cloud with the older tool version. Opus 5.5 accepts only the toolset there. You operate under zero data retention and need a model whose availability under it is documented — Opus 5's is, while the migration guide records nothing new for Opus 5.5 and says to treat it the same. Each is a reason to schedule the move, not a reason to skip it.

The reasons that are just inertia

"It works" is not a reason when the replacement is cheaper and the change set is known. Neither is a worry about losing reasoning mid-conversation: Opus 5.5 reads Opus 5's thinking blocks, so a conversation that moves onto it keeps its history. The reverse is where reasoning is lost — Opus 5.5's blocks are unreadable to Opus 5, which matters if a fallback routes back.

What to test before switching

Set effort explicitly — Opus 5.5 defaults one level lower — and compare outputs at matched settings, not at each model's default. Anthropic reports Opus 5.5 at medium beating Opus 5 at high on coding and knowledge work; that is the comparison to run on your own evaluation set.

Related

See migrating: Opus 5 to Opus 5.5 and claude-opus-5.