Claude Opus 5 Is Verbose by Design: Your Migration Guide

Claude Opus 5 launched on July 24, 2026, and within days developer forums were flooded with complaints: responses that would not end, verification passes nobody asked for, tasks that grew extra steps on their own. Dan Shipper called it "very hard to love." CodeRabbit noted "four times as many nitpicks." Token bills arrived that did not match the headline pricing.

The frustrating part: Opus 5 is doing exactly what Anthropic designed it to do.

This model is built for complex, long-horizon agentic work. It verifies its own output, narrates its progress, and proactively catches errors — all without being told to. The problem is that most prompt libraries were written for Opus 4.8, which needed to be told to do these things. When you ask Opus 5 to do something it already does, you get double the behavior, double the tokens, and a slower, more expensive experience.

The fix is not to fight the model. It is to stop asking it to do what it already does.

Why Lowering Effort Does Not Fix Verbosity

The first thing developers try is dropping the effort parameter. This is a reasonable instinct — lower effort should mean less work, right?

Not exactly. The effort parameter controls how much the model thinks (extended thinking tokens). It does not reliably shorten the visible response. To control response length, you have to prompt for it explicitly.

This is the single most common misconception about Opus 5.

The Six Behaviors to Retune

Anthropic's official Prompting Claude Opus 5 guide identifies six behavioral shifts that most often need tuning when migrating from Opus 4.8.

1. Response Length

Opus 5's default conversational responses run longer than Opus 4.8. Add an explicit conciseness instruction to your system prompt:

Keep responses focused, brief, and concise. Keep disclaimers and caveats short,
and spend most of the response on the main answer. When asked to explain
something, give a high-level summary unless an in-depth explanation is
specifically requested.

In longer system prompts, pair it with a short reminder near the bottom:

<tone_preference>
Keep outputs reasonably concise.
</tone_preference>

2. Agentic Narration

During agentic tasks, Opus 5 announces what it is about to do, narrates each step, and produces longer per-message output than prior models. To tune narration down to a signal-only cadence:

Before your first tool call, say in one sentence what you're about to do.
While working, give a brief update only when you find something important or
change direction. When you finish, lead with the outcome: your first sentence
should answer "what happened" or "what did you find," with supporting detail
after it for readers who want it.

3. Written Deliverable Length

Files that Opus 5 writes to disk — reports, Markdown documents, summaries — also run longer than prior models. Add explicit calibration if your product produces Claude-authored documents:

Match the length of written documents to what the task needs: cover the
substance, but do not pad with filler sections, redundant summaries,
or boilerplate.

4. Task Scope and Over-Verification

Opus 5 verifies its own work without being told to. If your prompt contains explicit verification steps ("include a final verification pass," "use a subagent to verify"), remove them. Those instructions compound with behavior the model already has. Removing them reduces wasted tokens with no quality loss.

The model also expands task scope on its own judgment. For narrow tasks, constrain scope explicitly:

Deliver what was asked, at the scope intended. Make routine judgment calls
yourself, and check in only when different readings of the request would lead
to materially different work. If the request seems mistaken or a better approach
exists, say so in a sentence and continue with the task as asked rather than
quietly narrowing, widening, or transforming it. Finish the whole task, and
stop short of actions that are clearly beyond what was asked.

5. Subagent Spawning

Opus 5 delegates to subagents more readily than prior models. Delegation is powerful for genuinely parallel work, but multiplies cost on small tasks. Give explicit guidance on when it is warranted:

Delegate to a subagent only for large tasks that are genuinely independent and
parallelizable, such as a wide multi-file investigation. Do not delegate work
you can finish yourself in a handful of tool calls, and do not use subagents
to verify or double-check your own work. If one subagent can complete the task,
use one rather than several, and keep spawn counts low.

6. Self-Correction Narration

Opus 5 catches and fixes its own mistakes without prompting. Avoid instructions like "double-check your answer" or "re-verify before responding" — these compound with built-in correction behavior and add cost without benefit.

The model also narrates corrections to earlier statements more than prior models do. To limit that narration to corrections that actually matter:

Only correct an earlier statement when the error would change the user's code,
conclusions, or decisions. State corrections plainly and briefly, then continue
the task. For slips that change nothing for the user, make the fix and move on
without noting it.

The Elephant in the Prompt Library

Boris Cherny's team at Anthropic removed more than 80% of Claude Code's existing system prompt when migrating to Opus 5. That is a dramatic number, and it is the clearest signal of what real migration work looks like.

Your existing prompt library was scaffolding built for a less capable model. Instructions like "verify your work," "think step by step," "check for errors before responding" were workarounds for Opus 4.8 behavior. Opus 5 does all of this unprompted. Keeping those instructions in place tells the model to do the thing twice.

The migration work is primarily subtraction, not addition.

Effort Levels Worth Knowing

Once you have retrimmed your prompts, effort levels become your primary cost lever:

• low and medium — Strong quality at a fraction of the tokens. Use these as your default and adjust upward based on evals.
• high — The default setting. A good starting point.
• xhigh — Reserved for demanding agentic coding and complex multi-file work where the quality difference justifies the cost.

If you carried effort defaults from Opus 4.8, re-run an effort sweep against your own evals. The recommended levels for Opus 5 differ significantly.

Quick Migration Checklist

Before going to production with Opus 5:

• Remove all explicit verification instructions ("verify," "double-check," "re-check")
• Remove step-by-step instructions — Opus 5 reasons by default when thinking is on
• Add explicit response length guidance if you need concise output
• Tune narration cadence for agentic tasks
• Audit subagent delegation rules — add explicit caps
• Run an effort sweep: default to low or medium, step up only where evals demand it
• If thinking is disabled: remove any "do not reason" instructions (they cause internal tag leakage)

The Bigger Picture

Anthropic's framing is instructive. Opus 5 is not a chatbot tuned for short, punchy answers. It is built for the kind of work where depth, verification, and multi-step planning add real value — complex codebase features, long-horizon agent tasks, enterprise document generation.

For simple tasks, low effort plus a tight conciseness prompt will get you most of the way there. For what Opus 5 was actually designed to do — difficult agentic coding, multi-agent coordination, large-scale document work — the verbosity is often doing useful work. The right tuning is about shaping where that work surfaces, not eliminating it.

The developers getting the best results from Opus 5 are not fighting its verbosity. They are redirecting it.