Every model page on the internet lists strengths. This one lists the other half, because knowing where Claude Opus 5 is weak is what stops you deploying it somewhere it will embarrass you.
Behavioural quirks you have to prompt around
- It writes long. Default responses, and files it writes to disk, run longer than earlier models. Lowering the effort level does not reliably fix it; only prompting does.
- It narrates. In agentic sessions it produces a lot of text between tool calls. Useful for oversight, noisy for a polished product.
- It can widen scope. Ask for a fix and you may get surrounding improvements you did not request.
- It delegates readily. With subagents available and no cap, an agent run can cost far more than you expected.
- It over-corrects itself. It sometimes narrates its own earlier mistakes at length, which reads as thrash in a user-facing product.
All five are prompt-tunable, which is the good news. All five will surprise you if you migrate a prompt library unchanged, which is the bad news.
Hard limitations
- No image, video, or audio generation. It analyses images; it does not create them.
- No Priority Tier, and a separate rate-limit bucket from the Opus 4.x models.
- Fast mode is Claude API only — not on Bedrock, Google Cloud, or Foundry.
- Server-side tool coverage varies by cloud platform. Web search, code execution, and batches are not uniformly available.
- Refusals happen. Elevated cybersecurity safeguards can decline a request, and benign security or life-sciences work occasionally trips them. Handle
stop_reasonbefore reading content.
Where it is the wrong tool
- High-volume simple tasks. Classification, extraction, and routing belong on Haiku 4.5 at a fifth of the input price.
- Hard latency budgets. Long turns are normal; if you need a sub-second answer, this is the wrong tier.
- Anything requiring guaranteed factual currency. Verify claims, dates, and numbers — always.
- The absolute frontier. Claude Fable 5 sits above the Opus tier for the hardest long-horizon work, at double the price.
The honest summary
Opus 5 is a strong default for hard work and a poor default for everything. The teams that get the most from it are the ones that route: cheap models for volume, Opus 5 for judgement, and a prompt library tuned for how this model actually behaves rather than how the last one did.
More on Claude Opus 5
Start with the complete Claude Opus 5 guide for the overview, or go deeper:
Ready to go deeper?
Read the full Claude Opus 5 guideFrequently Asked Questions
Can Claude Opus 5 generate images?
No. Claude analyses images — charts, screenshots, scanned documents, diagrams — at high resolution, but it does not generate them. For image generation you need a dedicated tool alongside it.
Why does Claude Opus 5 sometimes refuse requests?
It ships with elevated cybersecurity safeguards, and benign work adjacent to those domains — security tooling, some life-sciences tasks — occasionally trips them. A refusal returns a normal HTTP 200 with a refusal stop reason rather than an error, so check that field before reading the content. Server-side fallback can re-run declined requests on Opus 4.8 automatically.
Is Claude Opus 5 always the best Claude model to use?
No. Claude Fable 5 is more capable for the hardest long-horizon work, at double the price. Sonnet 5 and Haiku 4.5 are better choices for volume, latency, and cost. Opus 5 is the best default for hard work specifically, not for everything.
