Last updated: July 2026
Key Takeaways
- Opus 5 verifies its own work, expands scope, and narrates more than Opus 4.8, so instructions you wrote to compensate for older models now stack on top of behavior it already has.
- Lowering the effort setting will not reliably shorten Opus 5's responses. Anthropic's documentation states this directly: effort controls thinking volume, not visible response length.
- Anthropic published the fixes the same day it published the model, in a prompting guide most launch coverage skipped.
Claude Opus 5 shipped on July 24, 2026, and the reaction split within hours. The benchmark numbers are strong. The day-to-day experience, for a lot of people, is not: answers that will not end, corrections nobody asked for, tasks that grow extra steps, and token bills that do not match the half-price headline.
Nearly all of it is documented. Anthropic published a prompting guide for Opus 5 the same day it released the model, naming these behaviors directly along with what to change. The uncomfortable part: in four separate places it tells you to delete an instruction you wrote for Opus 4.8. If you have a mature prompt library, custom skills, or a harness you tuned over the past year, that library is now part of the problem.
What actually changed in Opus 5
Four behaviors shifted relative to Opus 4.8, and every complaint circulating today traces back to one of them. Opus 5's default responses run longer. It narrates more during agentic work, announcing what it is about to do. It verifies its own work without being told to. And it delegates to subagents more readily.
Each is defensible alone. The friction comes from collision: prompt libraries written for models that did not do these things are still telling Opus 5 to do them, and those instructions compound rather than override.
Anthropic frames this as ordinary tuning and says Opus 5 performs well out of the box on existing Opus 4.8 prompts. Early testers have been less generous. Lenny's Newsletter published a hands-on review the same day describing a neurotic personality in real coding sessions, including a merge conflict the model declined to touch. Treat that as one tester's experience, not a measured finding, though it matches what many are reporting.
It will not stop explaining itself
This is the most common complaint and it has the least intuitive fix. Anthropic confirms that Opus 5's default user-facing responses run longer than prior Opus models. The instinct is to turn the effort setting down. That does not work, and the documentation is explicit about why.
Effort governs how much the model thinks, not how much it says. On Opus 5 specifically, Anthropic's effort documentation states that changing the setting does not reliably shorten responses, and directs you to prompt for length instead. Anyone currently dragging the effort slider down to stop the verbosity is pulling a lever wired to something else.
The working fix is an explicit conciseness instruction in your system prompt: tell it to keep responses focused and brief, keep caveats short, spend the bulk of the response on the actual answer, and give a high-level summary unless depth was requested. In a long system prompt, repeat a one-line version near the end. Length instructions placed only at the top of a long prompt tend to get diluted.
Two related behaviors have separate fixes. Opus 5 narrates corrections to its own earlier statements more than previous models, which is the behavior people are calling neurotic. Constrain it by instructing the model to flag a correction only when the error would change your code, conclusions, or decisions, and to fix silent slips without announcing them. Separately, files it writes to disk run longer too, so if it is producing reports or Markdown documents, add a line telling it to match document length to the substance of the task rather than padding with summary sections.
It is doing too much, or too little
Scope creep and runaway subagents
Anthropic documents that Opus 5 can expand a task's scope, adding steps that were not requested and applying its own judgment about what the task should be. For narrow work, constrain that explicitly: instruct it to deliver what was asked at the scope intended, to make routine judgment calls itself, and to say so in a sentence if it thinks a better approach exists rather than quietly transforming the task.
The costlier version is delegation. Opus 5 spawns subagents more readily than prior models, which pays off on genuinely independent tracks of work and multiplies cost and time on small ones. Give it a rule: delegate only for large, parallelizable work, never to double-check its own output, and keep spawn counts low.
Claude Code users should know one thing that is easy to miss. The ultracode setting in Claude Code's effort menu is not a separate effort level. It pairs xhigh effort with standing permission for Claude Code to launch multiagent workflows. That combination, maximum exploration plus unprompted delegation, is precisely the configuration that produces the runaway sessions people are posting about. If your bill jumped today, check whether you are running ultracode.
Agent permissions are worth thinking about beyond cost, too. Our guide to secure AI agent alternatives covers the network-side controls that matter once an agent can spawn other agents on your machine.
Why it sometimes gives you less than you asked for
The opposite complaint, that Opus 5 stops early or under-delivers, has a specific documented cause. Opus 5 follows conservative instructions more literally than prior models. Anthropic's example is code review: a prompt saying to report only high-severity issues, or to be conservative, will produce genuinely less output on Opus 5 than the same prompt produced on 4.8. The model is not ignoring you. It is obeying an instruction you wrote when obedience was looser. The recommended pattern is to ask for everything and filter in a separate pass.
Your old prompts are the problem
Testers spent today discovering by trial and error that Opus 5 behaves better when you start from a clean slate and delete accumulated skills and prompts. Anthropic's documentation says the same thing, with reasoning attached. Verification instructions cause over-verification on Opus 5, and removing them cuts wasted tokens with no loss in quality. The same applies to legacy harness scaffolding that adds separate verification steps, and to re-check instructions like asking it to double-check before responding, which compound with behavior the model already performs.
Here is the full audit list. Everything in the middle column is something to delete, not rewrite.
| Symptom | Remove from your prompt | Add instead |
|---|---|---|
| Wasted tokens on redundant checking | Verification steps, subagent verification passes, harness scaffolding that verifies separately | Nothing. Opus 5 verifies without being asked. |
| Constant mid-answer self-corrections | Instructions to double-check or re-verify before responding | A rule limiting corrections to errors that would change your code, conclusions, or decisions |
| Responses too long | Nothing. This is not caused by an old instruction. | An explicit conciseness instruction, repeated briefly at the end of long prompts |
| Written files padded with filler | Nothing | A length-calibration line: cover the substance, skip boilerplate and redundant summaries |
| Task grows steps you did not ask for | Nothing | A scope rule: deliver what was asked, raise a better approach in one sentence, then proceed as asked |
| Cost spikes from too many subagents | Nothing, but check whether you are running ultracode in Claude Code | A delegation rule: large independent work only, never for self-verification, low spawn counts |
| Code review reports fewer issues than before | Conservative filters such as high-severity-only or be-conservative instructions | A request for everything, with filtering moved to a second pass |
| Internal tags leaking into visible output | Any rule telling the model not to think or not to reason | A general instruction to keep internal or system tags out of responses |
Guidance summarized from Anthropic's Opus 5 prompting and effort documentation, read July 24, 2026. Wording is ours; test against your own evaluations before deploying.
Two things to re-validate rather than delete. If you built prompt-side workarounds for weak vision handling on an older model, check whether they are still needed, because Opus 5 is materially stronger on charts, documents, and interface replication. And if you carried effort defaults across from a previous model, Anthropic recommends running a fresh effort sweep on your own evaluations rather than reusing them.
Which effort level should you actually use?
The advice forming in public today is to drop to low or medium effort. That is half right, and the half that is wrong will cost you on exactly the work you care most about.
Anthropic's position is that low and medium are genuinely stronger on Opus 5 than on earlier Opus models and should be used liberally as the primary control for token cost and response time wherever your evaluations show quality holds. But its recommended starting point for coding and agentic work is still xhigh, with high as the floor for most other intelligence-sensitive workloads. The API default is high, and xhigh only applies if you set it explicitly.
| Effort | Anthropic's guidance for Opus 5 | Watch for |
|---|---|---|
| low | Stronger here than on earlier Opus models. Simple tasks, high volume, subagent workers | Fewer tool calls and terser confirmations, which can read as skipped work |
| medium | Balanced. Use liberally wherever your evaluations show quality holds | It will not shorten visible responses. Prompt for that separately |
| high (default) | The floor for most intelligence-sensitive work. Applies if you set nothing | Many people are running here without knowing it |
| xhigh | Recommended starting point for coding and agentic work. Must be set explicitly | Thinking cannot be disabled here. Set a large output limit, around 64k as a start |
| max | Step up when a task justifies unconstrained token spending | Thinking cannot be disabled here either. Requests that try return an error |
Effort is a behavioral signal rather than a hard token budget, and its impact varies by task. Validate on your own workload before standardizing.
One trap that will catch people this week. Changing the effort level between requests invalidates your cached prompt prefix, because effort shapes the rendered prompt. If you follow the emerging advice and start switching effort per task inside one long session, you will lose your cache hits and your bill will rise even as you lower the setting. Vary effort across workloads, not within a conversation you are relying on caching for.
Strange XML tags or tool calls appearing as text
Both are documented artifacts of running Opus 5 with thinking disabled. The model can occasionally write a tool call into its visible text rather than emitting a real one, in which case the call never runs and the leaked text stays in conversation history, affecting later turns in an agentic loop. It can also emit internal tags into the visible response. Two counterintuitive details: a system-prompt rule telling the model not to think increases tag leakage rather than reducing it, and calling the tags out by name works less well than a general instruction to keep internal tags out of responses. Anthropic's primary recommendation is to leave thinking enabled and control cost through lower effort instead. On Opus 5 you cannot disable thinking at xhigh or max anyway; those requests return an error.
The one fix nobody can write for you
Every fix above is an instruction written against a model somebody else controls. Today the vendor told you to delete the instructions you wrote for the previous version. That will happen again on the next release, and the prompt library you rebuild this weekend has the same expiry date as the one you are deleting. This is the same measurement problem we worked through when testers claimed the returning Fable 5 had been quietly degraded, covered in our analysis of whether Claude Fable 5 was nerfed: when behavior and capability are bundled inside a service, you cannot tell them apart from outside.
The honest counterweight is that this tax is worth paying for genuinely hard work, and nothing you run at home matches Opus 5 on the benchmarks it just posted. But work that does not need frontier capability does not need this churn either. Our guide to the best local AI models by VRAM tier covers what runs at each memory level, the local AI hardware guide covers what each budget buys, and our zero-cost local agent stack walkthrough shows how to run the routine half of a workflow on hardware whose behavior changes only when you change it.
Frequently Asked Questions
Why is Claude Opus 5 so verbose?
Anthropic confirms that Opus 5's default user-facing responses run longer than prior Opus models, and that files it writes to disk are longer too. Neither is a bug or a broken setting. Both require an explicit length instruction in your prompt.
Does lowering the effort setting make responses shorter?
Not reliably on Opus 5. Anthropic's effort documentation states that effort controls thinking volume rather than visible response length, and directs you to prompt for length instead. Lowering effort will reduce thinking and tool calls, which cuts cost, but the visible answer can stay just as long.
Why is Opus 5 doing things I did not ask for?
Opus 5 expands task scope and applies its own judgment about what a task should include more than earlier models. It also delegates to subagents more readily. Both need explicit constraints: state the scope you want, and cap delegation to large independent work only.
Do I need to rewrite my prompts for Opus 5?
Mostly you need to delete parts of them. Verification instructions, re-check instructions, legacy harness verification steps, and rules telling the model not to think all now work against you. Anthropic says removing verification instructions cuts wasted tokens with no loss in quality.
Which effort level should I use for coding?
Anthropic recommends starting at xhigh for coding and agentic work, which you must set explicitly since the default is high. Step down to medium or low only where your own evaluations show quality holds. If you carried effort settings over from an earlier model, run a fresh sweep.
Why am I seeing thinking tags in Opus 5 output?
That artifact appears when thinking is disabled. Counterintuitively, a system-prompt rule instructing the model not to think makes it more likely, not less. The recommended fix is to keep thinking enabled and manage cost with lower effort instead.
