Claude Opus 5: worth the tokens on long documents, wasted on quick questions
A premium text model is not a better version of a cheap one at every job - it is a different bet about where mistakes cost you. Opus 5 is the bet worth making when the mistake would be buried on page forty of something nobody is going to reread.
- Cost
- 12 PT
- Per
- ~750 tokens
- Tier
- premium
- From
- $11/mo
Opus 5 earns its price on work where being wrong is expensive to discover: long documents that must stay consistent end to end, and code changes landing in a codebase someone else maintains. Most prompts are not that. For recall, reformatting and first drafts, a budget model reaches the same answer for a fraction of the tokens.
What Claude Opus 5 is best at
- Long documents that have to stay consistent end to end - a contract, a spec, a policy - where an answer that is right about page three and wrong about page forty is worse than no answer at all.
- Code changes that have to land inside an existing codebase and match its conventions, rather than arriving rewritten in the model’s own house style.
- Instructions with several constraints holding at once - tone, format, forbidden words, length - where the usual failure is quietly dropping the third one.
- Work you cannot audit yourself, where a fluent wrong answer is cheap to produce and expensive to discover later.
- Second-pass review of a cheaper draft: it reads a short input, writes a short output, and catches what the draft got away with.
Where it falls down
- Most prompts do not need it. On short, well-specified tasks a budget model gets there too, and paying flagship rates for work you could check at a glance is the most common way to burn a month’s budget in a week.
- You are billed on what you send as well as what comes back, so pasting a long document costs premium rate before a single word has been written.
- It writes thorough by default. Unconstrained answers run long, and you pay for the paragraphs you delete.
- Confidence is not accuracy. The price buys better judgement, not verifiability - a summary of a document you have not read still needs spot checks against the source.
- It is slower than the budget tier, which is felt most while you are still iterating on the prompt rather than running a finished one.
Prompts for Claude Opus 5
Copy one, swap the subject, keep the structure - the constraint clauses are what make these land.
Find the contradictions in a long document
Claude Opus 5Read the document below as one argument, not as separate sections. List every place where two parts contradict each other, quoting both and giving the section or page for each. Then list the claims that are asserted once and never supported. Do not summarise, do not suggest improvements, and answer "no contradictions found" if that is the honest result. The document follows this line.
Naming the failure mode you want found, and forbidding a summary, stops a long-context model from spending its answer restating the document back at you - and the explicit permission to find nothing is what makes the list worth trusting.
A refactor that respects the codebase it lands in
Claude Opus 5Before writing any code, list the files you would touch and a one-line reason for each, then stop and wait for me. Match the conventions already in the code - naming, error handling, test style - even where you would do it differently. If any part of my request conflicts with the existing design, flag the conflict instead of resolving it silently. The current implementation and the change I need follow this line.
Asking for a file-level plan and an explicit stop turns one expensive generation into a cheap one you can correct, which matters most on a model priced per unit of text it produces.
Grade a cheaper model’s draft instead of rewriting it
Claude Opus 5The text below was written by a faster, cheaper model. Do not rewrite it. Mark each paragraph keep, fix, or cut, and for every fix name the specific defect in one sentence - a claim that needs a source, a transition that hides a logical jump, a sentence that says nothing. Return only the marked list, no rewritten prose. The draft follows this line.
Critique reads a short input and writes a short output while generation does the opposite, so spending premium judgement on reviewing a draft costs far less than spending it on writing one.
Hold several constraints at once without dropping one
Claude Opus 5Write a 300-word update to customers about a two-hour outage. Every constraint must hold: no apology in the first sentence; never use the words "unfortunately" or "inconvenience"; name the root cause in plain language; state the specific change that stops it recurring; end on a factual sentence rather than a promise. After the draft, list each constraint and quote the part of your own text that satisfies it.
Making a model quote its own output back against each constraint converts a silent failure into a visible one, so you can see which rule it dropped without rereading the whole draft yourself.
Extract contract facts without inventing any
Claude Opus 5From the agreement below, extract: parties, effective date, term, renewal mechanism, termination rights for each side, payment amounts and due dates, liability caps, and governing law. Put the exact clause number next to every value. Where the agreement is silent on one of these, write "not stated" - do not fill it in from what is conventional. The agreement follows this line.
Requiring a clause number beside every extracted value makes invention expensive for the model and cheap for you to catch, because a wrong citation is far easier to spot than a wrong fact.
Craft notes
- 01Ask for the plan before the prose. An outline or a file list you can correct costs a fraction of the generation it prevents.
- 02Say how long the answer should be. Left unconstrained it writes thorough, and length is what you are billed on.
- 03Trim the input, not just the output. What you paste is metered too, so the three relevant sections beat the whole handbook.
- 04Keep follow-up questions in the same conversation. Repeated context bills at a fraction of the first read, and pasting the document into a fresh chat throws that discount away.
- 05Use negative instructions - "do not summarise", "do not rewrite", "return only the list". Being helpful in ways you did not ask for is the default, and you pay for the extra.
- 06Draft on DeepSeek V4 Flash and bring the result here for review. Buying premium judgement by the critique rather than by the draft is the single biggest saving available on this plan.
When to use something else
Claude Opus 5 is one of 26 text models on the same Platform Token budget, so switching costs nothing but the tokens.
The task is recall, reformatting, or a first draft you intend to rewrite anyway - anything you can check at a glance, where the cheapest tokens on the plan are genuinely enough.
You want the same instruction-following on ordinary drafting, extraction and code review without paying flagship rates for every message in a long back-and-forth.
You are working through a large pile of mixed material and throughput matters more to you than the last few percent of judgement on any single answer.
The answer will be audited and you want a second flagship from a different vendor on the same job, because two independent runs disagreeing is a cheaper error signal than finding the mistake downstream.
You are doing long-form work at volume and want a premium-tier answer without flagship pricing attached to every message in the thread.
Frequently asked questions
Holding a long document or a large codebase consistent in one pass. On short, well-specified tasks the gap narrows to almost nothing, which is why it should not be your default for everyday prompts.
Claude Opus 5, on the same bill as everything else
12 PT per ~750 tokens, shown before you spend it. Claude Opus 5 is on the $11/month plan and up.
Start free1 free prompt · no credit card · cancel any time