Anthropic publishes an official prompting guide for Claude Opus 5

The guide flags things that are easy to miss if you just carry old prompts over from prior models. Opus 5 verifies its own work without being told to — if a prompt still contains instructions like “add a final verification step,” it’s worth removing them, since they trigger redundant re-checking and burn tokens for no benefit. The same goes for “double-check your answer” style instructions — the model already catches its own mistakes reasonably well on its own.

Another detail: Opus 5 delegates to subagents more readily than earlier versions, which can meaningfully raise cost and runtime on small tasks unless you set explicit limits on delegation. The model also gives longer answers by default and narrates intermediate steps more often — if that’s not what you want, you have to explicitly ask it to be brief.

The guide points to a 1M token context window as both the default and the maximum, and Anthropic says instruction following, tool calling, and reasoning quality stay consistent across that entire window rather than degrading toward the tail end, which has been a known weak spot for long documents on earlier models.

Vision performance gets a specific callout too: chart, document, and diagram reading, plus UI and frontend visual replication, are strong enough that any prompt-side workarounds tuned for older models are worth re-testing, since they may no longer be necessary. The guide also notes that giving the model tools to crop, zoom, and visually check its own output works better than just relying on more thinking for vision-heavy tasks.

The effort parameter — low, medium, high, xhigh — is now the main lever for controlling response cost and speed. Notably, effort controls how much the model thinks, not how much it says, so trimming effort alone will not reliably shorten a chatty response — that needs an explicit conciseness instruction in the prompt instead.