claude-opus-5 — pricing, context window and limits

The previous Opus — still served, no longer the default

Opus 5 was the default recommendation for agentic and enterprise work until Opus 5.5 replaced it at a lower per-token price and a deeper cache-read discount. It is still served and still a sound model; the question now is whether anything keeps a workload here, and for most the answer is no.

The three reasons a team might stay

It is the last Opus that lets you switch thinking off, at effort high or below — Opus 5.5 rejects a disabled thinking configuration at every level. It still accepts forced tool choice, which Opus 5.5 rejects. And its availability under zero data retention is documented, while the migration guide records nothing new on retention for Opus 5.5 — confirm with your account team before moving a zero-retention workload. If none of those apply, staying mostly means paying more per token for a model Anthropic's own testing places below its replacement.

The traps it introduced still apply here

Thinking is on by default on this model, which changed several prior Opus generations' behaviour and caught out code that sized max_tokens for answers alone. Disabling thinking above high effort is rejected — see thinking disabled above high effort on Opus 5. Its minimum cacheable prefix is lower than Opus 4.8's, so prompts too short to cache there may cache here with no code change.

Fast mode is available, at its own rate

Where latency matters more than standard-speed pricing, fast mode applies at its own premium rate — see how fast mode changes the price of Opus.

Related

See migrating: Opus 5 to Opus 5.5 for the move, and Opus 5.5 vs Opus 5 for when staying still makes sense.