Verdict: Claude Opus 5 (released July 24, 2026) makes most of the old prompt-engineering rituals not just unnecessary but actively harmful. Anthropic's own model-specific prompting guide says to delete "check your work" instructions, stop drip-feeding tasks, ask for all issues (not just serious ones) in reviews, and let the effort dial — not your prompt wording — control quality. What survives is the boring half: clear context, real constraints, and a decent brief. If you can brief a person, you can prompt Opus 5.
Last verified: 2026-07-31 · Pricing/limits change often — last checked July 31, 2026.
- Opus 5 costs $5/M input, $25/M output — same as Opus 4.8, half of Fable 5 (Anthropic pricing docs)
- Thinking is ON by default; disabling it at xhigh/max effort returns a 400 error
- Effort (not prompt wording) is now the primary quality/cost lever
- Four things to delete from your prompts + two constraints to add + one rule for reviews
What changed with Claude Opus 5 that makes old prompts worse?
Claude Opus 5 verifies its own work unprompted, writes longer responses by default, delegates to subagents more readily, and narrates progress more often than Opus 4.8 did. These are all things the old guard of prompt engineering spent years trying to force models to do — with instructions like "double-check your work," "include a verification step," and "be thorough."
On Opus 5, those instructions compound with behavior the model already has. You don't get double the accuracy; you get double the verification passes, double the tokens, and double the latency — for zero quality gain. Anthropic's official prompting guide puts it bluntly: remove explicit verification instructions because they cause over-verification and waste tokens with no improvement when removed. (Prompting Claude Opus 5 — Anthropic Docs)
The block below summarizes the shift. The old playbook was additive — pile on instructions to push the model harder. The new playbook is subtractive: strip out the rituals that are now built in, then add back only the constraints the model's new tendencies require.
| What you used to add | What to do on Opus 5 | Why |
|---|---|---|
| "Check your work before responding" | Delete it | Opus 5 self-verifies by default; adding it causes over-verification |
| "Break this into steps, one at a time" | Give the full spec upfront | Opus 5 does its best work from one complete brief, left to run |
| "Be thorough and complete" (no scope cap) | Write explicit scope limits | Opus 5 expands narrow tasks on its own — extra work costs your time and tokens |
| "Only flag serious issues" (for reviews) | Ask for every issue | The model follows "be conservative" literally and suppresses findings you need |
| "Set thinking to max effort" (carried from 4.8) | Start at high, sweep down | Low and medium effort are now meaningfully stronger than on prior models |
| Role-play framing ("You are a senior…") | Keep only for genuine voice | A role line still works, but elaborate personas are no longer needed |
How does the effort setting work on Claude Opus 5?
Effort is the primary control for trading off intelligence, speed, and cost on Opus 5 — and it matters more than any single line in your prompt. The effort parameter controls how much the model thinks before responding, and Opus 5 converts extra effort into better results more reliably than any earlier Opus model. (Effort — Anthropic Docs)
The five effort levels are:
| Effort | Use it for | Token/latency cost |
|---|---|---|
| Low | Simpler tasks, subagents, high-volume classification | Lowest — "significant token savings" per Anthropic |
| Medium | Balanced speed/cost/performance for everyday tasks | Low — strong quality at a fraction of tokens |
| High (default) | Complex reasoning, coding, agentic tasks | Standard — same as omitting the parameter entirely |
| Xhigh | Hard coding and agentic work that needs deeper reasoning | High |
| Max | Unconstrained, deepest reasoning for hardest tasks | Highest — set large max_tokens (64k+) |
Anthropic specifically calls out that low and medium effort on Opus 5 produce strong quality at a fraction of the tokens and latency compared to prior models. If you carried effort settings over from Opus 4.8 or 4.7, run a fresh effort sweep — the levels were recalibrated. (Prompting Claude Opus 5 — Anthropic Docs)
Key distinction: effort controls thinking volume, not response length. Lowering effort does not reliably shorten visible responses. If you want shorter answers, you have to prompt for brevity explicitly.
Thinking is on by default — and that's a breaking change
On Opus 4.8, a request without a thinking field ran without thinking. On Opus 5, the same request runs with thinking — the model decides when and how much to think. The one breaking change: thinking: {"type": "disabled"} is accepted only when effort is high or below. Pair it with xhigh or max and you get a 400 error. (What's new in Claude Opus 5 — Anthropic Docs)
Since max_tokens is a hard cap on total output (thinking + response text), a ceiling tuned tightly for a non-thinking workload on 4.8 can truncate mid-answer on Opus 5. Anthropic recommends starting at 64k max_tokens when running at xhigh or max effort.
What should you delete from your Opus 5 prompts?
Search your system prompts, CLAUDE.md files, and saved prompts for these patterns and remove them. Anthropic's guide is explicit: these instructions cause over-verification — the model runs a review it already ran, over and over again, burning tokens and time with no quality improvement.
Delete these verification patterns:
- "Double-check your work before responding"
- "Verify each step before moving to the next one"
- "Review your answer for errors, then revise it"
- "Include a final verification step for non-trivial tasks"
- "Use a subagent to verify"
- "Make sure the output is correct before returning it"
The exception: if you have genuinely high-stakes steps where an explicit check adds value, scope the instruction to that step only rather than making it a global rule. For example:
"Do not add general verification passes; you already verify by default. The only exception: after writing the migration SQL, run it against the schema dump once and report any mismatch."
You can still create a separate review agent — spawning a separate subtask with fresh context to review output is still useful and works natively. What's dead is the inline "check your work" bolted onto the end of every prompt. For more on multi-agent verification patterns, see our deep dive on building verification skills for multi-agent workflows in Claude Code.
How should you write prompts for Opus 5? (The four-block method)
The replacement for the old ritual incantations is a simple four-block structure. Each block addresses a specific Opus 5 behavior that needs managing — not a generic "best practice."
Block 1: The complete brief (replaces step-by-step feeding)
For about two years, the advice was to break jobs into steps and feed them one at a time — analyze, then plan, then build — because older models drifted on long jobs. Opus 5 does its best work from one complete brief handed over at the start and left alone to run. Open with the entire assignment: who it's for, what to build, what materials to use, and where the finished thing goes.
Open with the whole job in a single message:
"Build a one-page site for my community that does [X]. Everything you can use — the numbers, the description, what's inside — is on this page: [link]. Deliver the finished page as an artifact."
Block 2: Scope constraints (prevents scope expansion)
Opus 5 expands narrow tasks on its own — you give it a page and it offers a blog. Extra work sounds like a gift until it's your time reviewing things you never wanted and your usage limits paying for them. Write the limits down:
"One page only. No blog, no pricing tiers, no email capture form. Keep every number exactly as it appears. Do not invent testimonials, results, or member counts."
Anthropic's official scope-control block is the most versatile single addition:
"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. Finish the whole task, and stop short of actions that are clearly beyond what was asked."
Block 3: Output caps (controls verbosity)
Opus 5 overwrites in two places: the chat reply and the artifact it builds. One cap doesn't cover both. Put a number on the artifact — "five sections, no more" — and a separate cap on the reply: "When you're done, reply with two lines: a link to the page and anything you changed from the original."
For general conciseness, Anthropic provides this block:
"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."
And for written deliverables (reports, documents, summaries written to disk), which bloat independently of conversational verbosity:
"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."
Block 4: Cosmetic autonomy (reduces unnecessary check-ins)
Tell the model to make small decisions itself and keep going. Only ask you when a decision genuinely requires your input:
"Anything cosmetic — font, spacing, colors — decide on your own and keep going. If a decision genuinely needs my input, you can ask. Otherwise make the call."
How should you review work with Opus 5? (The counterintuitive rule)
The last rule is for reviewing, not building — and it runs opposite to the instinct most people have. When you hand the model something already finished (a landing page, a contract, a sales email) and ask it to review, most people say "only flag the serious issues" to avoid a list of 40 trivial notes.
You think you're telling it how carefully to look. What actually happens: it looks, finds everything, then filters before handing you a short list — and you never find out what was on the long list. Anthropic's guide says to flip this entirely:
"List every issue that you find, big or small. I will decide what matters."
You do the filtering yourself after seeing everything the model found. The model is better at finding issues than at guessing which ones you'll care about. This principle also applies to code review: Opus 5 finds real bugs at a high rate with few false positives, but if your prompt says "only report high-severity issues," it reports less — following the instruction literally and suppressing findings you need. (Prompting Claude Opus 5 — Anthropic Docs)
For a structured approach to this two-pass review pattern, our guide on building an AI follow-up cadence with an approval gate covers how to separate detection from decision-making in practice.
How much does Claude Opus 5 cost?
Claude Opus 5 is priced at $5 per million input tokens and $25 per million output tokens — unchanged from Opus 4.8 and exactly half of Claude Fable 5's $10/$50 rate. The 1M token context window is included at no premium. (What's new in Claude Opus 5 — Anthropic Docs)
| Token category | Price per million tokens |
|---|---|
| Input (standard) | $5.00 |
| Output (standard, includes thinking) | $25.00 |
| Prompt cache write (5-min TTL) | $6.25 |
| Prompt cache write (1-hour TTL) | $10.00 |
| Cache hit | $0.50 |
| Batch API input / output | $2.50 / $12.50 |
| Fast mode input / output | $10.00 / $50.00 |
Your bill can still rise even at the same per-token rate, because thinking is on by default and thinking tokens bill as output ($25/M). Lowering the effort setting — or stacking cache and batch — is the first lever to pull. For enterprise-level cost optimization strategies, see our guide on AI token cost optimization for enterprises.
Which plan should you use Opus 5 on?
On the Claude application (not the API), the model and effort settings you pick matter more than any single line in your prompt. Here's the practical tier guidance:
| Plan | Model recommendation | Effort guidance |
|---|---|---|
| Pro ($20/mo) | Sonnet 5 for daily work, switch to Opus 5 when the job matters | Medium for routine, bump to high for harder jobs |
| Max ($100 or $200/mo) | Opus 5 as the daily driver — it's roughly half the price of Fable 5 in token terms | Medium as default; high for complex work; xhigh for demanding agentic tasks |
| API | claude-opus-5 for production |
Start at high (the default), sweep low/medium on your evals |
On Max, stop opening the model picker — Opus 5 is the default, and it's well-suited as a daily driver. The effort dial is where you make the real trade-off. For a deeper comparison of when Opus 5 is worth the premium over Sonnet, see our GPT-5.6 Sol vs Claude Opus 5 comparison.
What survives from old prompt engineering?
Not everything died. The boring half survived and actually got more important:
- Clear instructions — state what you want, not what you don't want (positive examples beat negative instructions on Opus 5).
- Real context — who the work is for, where it goes, what matters. Opus 5 can only use the context you give it.
- Why the task matters — the model calibrates effort and scope better when it understands the purpose.
- An example when you need a specific format — if you want JSON, show the shape you want.
- Role-play for genuine voice — a simple role line ("write as a tax accountant") is still useful when the output voice genuinely matters. But elaborate persona constructs are no longer needed.
The skill that actually lasts isn't memorizing prompts — it's knowing that model-specific prompting guides exist and checking the one for the model you're on. Anthropic keeps one per model. After every model update, re-read it before you rewrite all your saved prompts, because what helped yesterday can hurt tomorrow. For a systematic framework on adapting your AI workflows to model changes, see our guide on building a model-agnostic AI agent OS that survives model churn.
What this means for you
If you're a business owner, developer, or anyone using Claude for real work, the takeaway is simple: audit your saved prompts and delete the ritual half. Look for "check your work," "be thorough," "verify each step," and "only flag serious issues" — those lines were written for a model that needed the help. On Opus 5, they're either dead weight or actively costing you tokens and time.
Then add back the four blocks: a complete brief, explicit scope limits, separate output caps for the chat and the artifact, and cosmetic autonomy. Set the effort dial to medium for daily work and bump it when the job genuinely needs more. For reviews, ask for everything and filter yourself.
The result: faster, cheaper, better output from the same model — not because the model got smarter, but because you stopped getting in its way.
FAQ
Q: Does prompt engineering still matter with Claude Opus 5?
A: Yes, but the shape changed. The "ritual" half — verification instructions, step-by-step feeding, effort proxies in prompt wording — is dead or harmful. What matters now is writing a clear brief with real context, explicit scope limits, and output caps. Anthropic calls this "context engineering" rather than prompt engineering.
Q: Should I delete "check your work" from all my prompts?
A: Yes. Anthropic's official guide says Opus 5 self-verifies without being told. Adding verification instructions causes over-verification — the model runs checks it already ran, burning tokens and latency for no quality gain. The one exception: if a specific high-stakes step needs an explicit check, scope the instruction to that step only, not as a global rule.
Q: What effort level should I use on Opus 5?
A: Start at high (the default). Use medium for everyday tasks — Anthropic says it delivers strong quality at a fraction of the tokens. Step up to xhigh for demanding coding or agentic work, and max only when a task justifies unconstrained token spending. If you carried effort settings from 4.8 or 4.7, run a fresh effort sweep — the levels were recalibrated.
Q: Can I turn off thinking on Opus 5?
A: On the Claude API, yes — but only at effort high or below. Setting thinking: {"type": "disabled"} with xhigh or max effort returns a 400 error. With thinking disabled, Opus 5 can occasionally leak tool calls into text output or XML tags into visible responses. The recommended path is to keep thinking on and control cost with lower effort instead.
Q: How much does Claude Opus 5 cost?
A: $5 per million input tokens and $25 per million output tokens — the same rate as Opus 4.8 and exactly half of Fable 5's $10/$50. Prompt cache hits cost $0.50/M, batch processing cuts both directions 50% to $2.50/$12.50, and fast mode runs $10/$50 for 2.5x speed. The 1M token context window is included at no premium.
Q: Should I use Opus 5 or Sonnet 5?
A: On Max plans, Opus 5 is the default and worth it as a daily driver. On Pro ($20/mo), use Sonnet 5 for daily work and switch to Opus 5 when the job matters. On the API, Sonnet 5 is roughly 40% of the cost ($2/$10 through August 2026, rising to $3/$15 after) and wins on high-volume classification and drafting; Opus 5 earns the gap on long agentic runs where fewer retries beat a cheaper per-token rate.
Q: What happens if I feed Opus 5 tasks one step at a time?
A: It still works — it's just no longer the default or the best approach. Opus 5 does its best work from one complete brief handed over at the start and left alone to run. Drip-feeding steps was a workaround for older models that drifted on long jobs; Opus 5 stays on task across extended tool-use sessions without the short leash.
Discussion
0 comments