Claude Opus 5: the model that tells you what you didn't ask

Anthropic's headline numbers for Opus 5 are about agentic coding. This isn't that. I wanted to know what the new model does with ordinary office work, the sort most people actually hand it, so I ran seven tasks against four Claude models: Opus 5, the Opus 4.8 it replaces at the same per-token price, Sonnet 5 and Haiku 4.5. Three runs each, two prompt versions per task, 168 API calls, identical settings throughout. Every input is synthetic and built for the test, so no client data went anywhere near it. Outputs were scored blind and unblinded afterwards.

Every task was run twice, once with a carefully written prompt, and once with the sort of one-liner a busy person actually types.

Here's what I found. On everyday drafting, summarising, rewriting, turning notes into a procedure, all four models produced usable work, and the differences between them were stylistic. Opus 5's real advantage is something else entirely: it tells you about problems you didn't ask it to look for.

The test that actually separated them

I built a sales table with a deliberate error in it. The East row is shown as 235, but its quarters are 60 + 70 + 85, which is 215. The grand total of 1,200 is correct, so the mistake sits in a single cell, exactly like the ones that survive three rounds of review because everybody's reading the bottom line.

Then I gave it the lazy prompt, which is the realistic one: "give me three takeaways from this for a management update". Nothing in that asks anyone to check anything.

Opus 5 gave the three takeaways, then added a section nobody asked for. All three runs did it:

"One correction before you circulate: the East row total is wrong, 60 + 70 + 85 = 215, not 235. That also means the Total column adds to 1,220, not 1,200. The quarterly column totals and the 1,200 grand total are correct, so the fix is just the East figure."

"One correction before you circulate: East's row total is shown as 235 but the quarters sum to 215 (60+70+85). The grand total of 1,200 is correct, so it's the East total that needs fixing."

"One data check before you circulate: the East row total is shown as 235, but 60 + 70 + 85 = 215. The column totals and the 1,200 grand total are internally consistent, so the East figure looks like the error, but the row totals sum to 1,220, so please confirm which figure is right before the numbers go out."

That third one is the best answer anything gave me in the entire test. It finds the error, works out which of the two candidate figures is more likely to be the wrong one, and still tells you to confirm before the numbers leave the building. It's what you'd want from a competent colleague who noticed something on their way past.

The other three built on the wrong number

They didn't quietly omit the error. They built on it.

Opus 4.8:

"North leads, East and West lag. North is the standout performer at 405k units (34% of total), while East (235k) and West (250k) together account for just 40%, suggesting scope to review strategy or resourcing in the weaker regions."

Haiku 4.5:

"East region underperforming: East region sales lag significantly at only 235k units (20% of total), with the slowest quarter-on-quarter growth rate. This region warrants attention to identify improvement opportunities."

Sonnet 5:

"East remains the smallest region but is closing the gap. Despite having the lowest total (235k units), East posted the highest percentage growth of any region."

Every one of those is a recommendation to go and look at a region's performance, generated from a number somebody typed in wrong. That's the everyday version of this problem, and it's why it matters: a colleague sends you a spreadsheet, you ask for a summary, and you act on it.

Tell it to check and the gap mostly closes

With the careful prompt, the one ending "flag anything that looks wrong or does not add up":

Sonnet 5 is every bit as good as Opus 5 once it knows what it's looking for, and Haiku couldn't do it either way. What separates the tiers here is initiative rather than capability, which matters more than it sounds, because if you always write careful prompts you can happily buy the cheaper model. Almost nobody always writes careful prompts.

It corrects a false summary too

Second task, same shape. Support metrics that get worse across a quarter, response time sliding from 4.2 to 6.8 hours and CSAT from 88% to 79%, with a note from ops attached claiming "Great quarter. Response times and satisfaction both trending up."

Opus 5 caught the contradiction in 3 of 3 runs on both prompt versions, and it was the only model that kept the update and the correction separate instead of blending them into one muddy paragraph:

"One flag: the ops note says 'response times and satisfaction both trending up,' but the data shows response times getting slower and CSAT falling. Worth clarifying with ops before this goes out, in case they were reading the response-time figure as an improvement."

The clause about ops possibly reading the response-time figure as an improvement is the bit I'd want from a person, because it works out why somebody made the mistake instead of just being right at them.

Sonnet 5 scored 0 of 3 on the careful prompt here, which sounds worse than it is. It reported the decline accurately every single time. It just never pointed out that the note attached to the data said the opposite.

Everything refuses the impossible ask now

Third task is unanswerable on purpose: two lines of vendor information, and an instruction to name the winner and give a three-year cost saving.

All four models declined, on both prompt versions, in 23 of 24 runs. Not one of them invented a number.

This used to be what separated a good model from a bad one, and it doesn't any more. Worth saying plainly, because it's good news and because you can check it yourself in about a minute. Every model in the line-up clears it, so none of the credit here belongs to Opus 5.

What it costs

Real spend on real API calls, 42 per model, computed from actual token counts.

Opus 5 and Opus 4.8 cost exactly the same per token. The 59% difference in spend is entirely because Opus 5 writes 73% more, and it's also the slowest of the four by some distance.

It has one genuinely annoying habit: it overshoots explicit length limits. Median output tokens where the prompt set one:

Haiku is the only model in the set that does what it's told on length. If brevity is the actual job, the expensive model is the wrong tool, and being cleverer doesn't help.

What I'd actually do with it

Use the cheap tiers when you already know what you're looking for. Haiku handles routine summarising and rewriting at roughly a thirteenth of the price and three times the speed, and it's the only one that respects a word count. For work where you've already done the thinking and you just want it typed up properly, paying more buys you a longer wait and a longer answer.

Use Opus 5 when you're handing it something you haven't checked yourself. A spreadsheet from a colleague, a report you're about to forward, figures going into a board pack. It's the only model in the ladder that raised the problem I hadn't asked about, and it did it every time. That costs about 1.7 cents a task, a bit over a penny, which is easy enough to justify for anything with your name on it that other people will act on.

The uncomfortable part is that Opus 5 isn't seeing something the others can't. All three of them found the same error the moment I asked them to look. What the extra money buys is a model that mentions it anyway.

What this doesn't tell you

Seven tasks, three runs each, scored by one person. That's a structured test rather than a study, and you should treat it as one.

It's Claude only. No GPT, no Gemini, so it says nothing about how Opus 5 stacks up outside Anthropic. Every input is synthetic and built for this purpose, including the sales table, so nothing here came from a real client document. And none of it tests Anthropic's own benchmark claims, which are about agentic coding rather than office work.

For the record, I pay for Claude Max and my agents run on Claude, the same disclosure I made on the Fable 5 piece. Make of the enthusiasm what you will, though on this evidence I'd still tell you to run most of your work through Haiku.

FAQ

What is Claude Opus 5 good at?

Noticing things you didn't ask it to notice. In a controlled test it flagged a planted arithmetic error in a sales table in 3 runs out of 3 when the prompt only asked for a summary. The other three Claude models flagged it in 0 of 3, and all three then wrote management recommendations built on the wrong number.

Is Claude Opus 5 better than Opus 4.8?

On ordinary drafting, not noticeably. On catching errors you didn't ask about, substantially. It's also 44% slower and cost about 59% more across the same 42 calls, despite identical per-token pricing, because it writes considerably more.

Do I need the most expensive Claude model?

For routine summarising and rewriting, no. Haiku 4.5 did those tasks at about a thirteenth of the cost and three times the speed, and it was the only one that respected a word limit. Save the expensive tier for documents you haven't checked yourself.

Does Claude make things up?

In this test, no. Asked to recommend a vendor and project a three-year saving from two lines of information, all four declined rather than inventing a number, in 23 of 24 runs. Not one produced a figure.

How was this tested?

Seven work tasks, four Claude models, three runs each, two prompt versions per task, 168 API calls in total with identical settings. Inputs were synthetic. Outputs were scored blind, before the model labels were revealed.