Claude Opus 5, explained: what it is, how it works, and how it ranks

Anthropic released Claude Opus 5 on 24 July. Here's what it is, how it works, where it sits against everything else, and the one part that actually matters if you're running a business in the UK.

What it is

Opus 5 sits in the middle of Anthropic's range. Above the small fast models, below Fable 5, which is the expensive one.

What it's for is the part people skip. A chatbot answers you in a few seconds. Opus 5 is built for the opposite job: long stretches of work with nobody watching. Reading two hundred supplier contracts and pulling out the three that don't match the others. Rebuilding a spreadsheet someone left behind. Making a software change that touches forty files. Driving a piece of software by clicking round it, the way a person would.

Anthropic call this "agentic". The honest word is unsupervised, and it comes back at the end of this piece.

How it works

It thinks first

Older models started typing straight away. This one reasons privately first, then writes, and that reasoning is switched on by default rather than something you turn on. It's the difference between the colleague who answers instantly and the one who goes quiet for a second and then says something useful.

You pay for those thinking tokens, which is why the next bit exists.

There's a dial for how hard it tries

Five settings per request: low, medium, high, xhigh, max. High is where it starts. Low and medium are much cheaper and faster, and on routine work the quality usually holds up fine. Max is for the problem you'd otherwise hand your best person and lose them for a day.

Most firms wasting money on AI are running everything flat out, including the boring 80%. Turning the dial down on the routine stuff is most of the bill.

It holds about 700,000 words at once

A million tokens, and a token is roughly three-quarters of a word. Call it a few thousand pages held in working memory.

Here's why that's worth caring about. Tools with smaller windows read long documents in pieces, so they'd miss the contradiction between page 4 and page 380, because they never had both in front of them. That failure mode mostly goes away at this size. Anthropic also say accuracy holds up across the whole window rather than tailing off at the far end, which was always the catch before.

It checks its own work

This is the one with the longest shadow. Sitting in the developer notes is an instruction to delete the verification steps engineers built into their systems. The "check your answer before you respond" scaffolding that's been standard practice for three years now makes the model over-verify, because it already does it.

A vendor telling you its product has outgrown the guard rails you built round it is not a small thing to put in a footnote. It also writes longer answers now, talks through what it's doing more, and hands work off to sub-agents more readily.

How it ranks

Benchmarks get picked by the people announcing them and none of them is your actual work, so treat all this as a rough map.

Frontier-Bench v0.1 is 74 tasks testing whether a model can finish real multi-step engineering jobs in a terminal instead of talking about them. Closest thing we've got to a test of working unsupervised. Scores at maximum effort:

• Claude Opus 5: 43.3%
• GPT-5.6 Sol (OpenAI): 37.5%
• Claude Fable 5: 33.7%
• Claude Opus 4.8, the one it replaces: 18.7%

43% sounds unimpressive until you notice it more than doubled its predecessor in one release, on exactly the category of work that automates jobs.

On SWE-bench Verified, which uses real bugs from real open-source projects, it scores 96%. That's the highest anyone has recorded, above Mythos 5 at 95.5% and Fable 5 at 95%. On the harder SWE-bench Pro it gets 79.2%, and on OSWorld 2.0, which is about driving a computer through its actual interface, 70.6%.

Against the rest of the field it leads GPT-5.6 Sol on Frontier-Bench, ARC-AGI-3, GDPval, OSWorld and AutomationBench. GPT-5.6 Sol takes DeepSWE and stays competitive on agentic search. Google's Gemini 3.1 Pro still holds the largest context window going at 2 million tokens, and the best scores on video and image work.

The useful takeaway isn't who won. It's that the gap at the top is the narrowest it's ever been, so what decides your choice now is price, availability and fit, not raw capability.

The bit that's specific to the UK

Look at that SWE-bench list again. Second place is Mythos 5, and you can't use it.

Back in June, as we wrote in our piece on the Fable 5 ban, a US export-control directive took Fable 5 and Mythos 5 offline worldwide for nearly three weeks. Fable 5 came back on 1 July. Mythos 5 is still restricted to approved US organisations. If you're a firm in Guildford, the most capable model on that leaderboard isn't on your menu.

Which makes this release unusually relevant here. The best model a UK business can actually buy is now also the mid-priced one, available on the API, AWS, Google Cloud and Microsoft Foundry. Capability, price and access haven't lined up like that before.

What it costs

$5 per million tokens in, $25 out. Identical to the model it replaces, half of Fable 5. Batch processing halves both again. Anything it's already seen drops to $0.50 per million. There's a Fast mode at double the price for about 2.5 times the speed, worth it when someone's sitting there waiting.

Run a long contract through it and you've spent a fraction of a penny. The number only gets real at volume, and that's the case worth modelling, because work you costed in 2025 and shelved as too expensive may now sit the other side of the line. The price per unit of capability has roughly halved twice since then. Document review, reconciliation, first-pass drafting. Our document intelligence demo is the fastest way to see what current models do with your own paperwork.

Why a better model makes your paperwork harder

Back to that word. Unsupervised.

Every step like this shifts work from "a person did it with AI help" towards "AI did it and a person approved it". The output improves. The audit trail thins out. Both happen at once and only one of them is visible from the outside.

Regulators here have already moved. The FCA says it'll publish a good and poor practice report on AI in financial services later this year. The Data (Use and Access) Act 2025 turned automated decisions from a near-prohibition into a right of challenge, with safeguards you have to be able to evidence. The SRA, the ICO and the CQC are all asking their own version of the same question.

None of them asks which model you used. They ask who was accountable, what data it touched, and how the output got checked. A better model answers none of that, and arguably makes it harder, because the more the machine does on its own the more of your defence rests on showing a human properly looked at it. Our guide to AI audit readiness covers the four documents that count as proof. Shadow AI covers the thing we find first in nearly every audit, which is staff already using tools nobody signed off.

What to actually do

Don't rebuild anything. Fourth capable frontier model in four months, and one of them got switched off by a government mid-quarter. If swapping model means re-engineering your process, your process is the problem.

Re-cost what you shelved. The economics moved while you weren't looking. That's the real opportunity in this release.

Sort out where your data goes before you scale. Capability is cheap and increasingly interchangeable. Data handling isn't. If client material can't go near a public API, a private deployment is far cheaper to plan than to retrofit.

Keep the human check and write it down. The better the model gets the more tempting it is to stop looking, which is exactly when that review step stops being a formality. Our compliance agent shows the pattern: machine does the legwork, a named person owns the outcome.

The bottom line

Opus 5 is a real step up and the interesting part isn't the scores. It's that near-frontier capability now costs half what it did six weeks ago, and the models above it are either twice the price or off-limits to you.

Which points where it always points. The model is a commodity that keeps leaping ahead, and your advantage was never owning the best one. It's having your data organised, your workflows mapped and your governance written down, so each leap is something you inherit instead of a project you have to run.

The firms that get the most out of this one are the ones that did that work before it existed.