Anthropic released Claude Opus 4.8 on 28 May 2026. It is their most capable model, and it now sits inside Claude and Cowork. If you use Claude for work, you are already on it. Here is what is new, in plain English, and the part most of the coverage will skip: a better model does not fix how Ai gets used inside a business.
The short version
Opus 4.8 is an upgrade on Opus 4.7. Anthropic describes it as a modest but tangible improvement, not a reinvention. It scores higher across Anthropic’s benchmarks for coding, reasoning, and knowledge work, and it runs at the same price as the version before it. For most people the day-to-day difference is two things: it is more honest about its own work, and you now get a dial that controls how hard it thinks.
How it compares
Anthropic’s own benchmark figures, set against Opus 4.7 and the latest models from OpenAI and Google. Opus 4.8 leads on five of the six. GPT-5.5 still wins on terminal coding, which is worth being straight about.
| Benchmark | Opus 4.8 | Opus 4.7 | GPT-5.5 | Gemini 3.1 Pro |
|---|---|---|---|---|
| Agentic codingSWE-Bench Pro | 69.2% | 64.3% | 58.6% | 54.2% |
| Agentic terminal codingTerminal-Bench 2.1 | 74.6% | 66.1% | 78.2% | 70.3% |
| Multidisciplinary reasoningHumanity’s Last Exam, with tools | 57.9% | 54.7% | 52.2% | 51.4% |
| Agentic computer useOSWorld-Verified | 83.4% | 82.8% | 78.7% | 76.2% |
| Knowledge workGDPval-AA score | 1890 | 1753 | 1769 | 1314 |
| Agentic financial analysisFinance Agent v2 | 53.9% | 51.5% | 51.8% | 43.0% |
Source: Anthropic, Introducing Claude Opus 4.8. GDPval-AA is a score, not a percentage. Reasoning row shows the with-tools result.
Watch
It owns up when it is not sure
The change worth your attention is honesty. Ai models have a habit of sounding confident when they should not. They claim a task is done, or that code works, when the evidence is thin. That is the behaviour that quietly costs you time, because you trust an answer that was never checked.
Anthropic trained Opus 4.8 to flag uncertainty instead of papering over it. In their own evaluations the model is around four times less likely than Opus 4.7 to let a flaw in code it has written pass without comment. It also scores far lower on Anthropic’s measure of misaligned behaviour, close to Mythos Preview, their best-aligned model.
Misaligned behaviour
Score 1 to 10 · lower is better
Rebuilt from Anthropic’s alignment chart. Opus 4.8 scores far lower than Opus 4.7 and lands close to Mythos Preview, Anthropic’s best-aligned model. Figures read from the published chart and are approximate.
For a business that matters more than any benchmark. The risk with Ai was never that it could not write the email or the report. It was that it would write the wrong one with total confidence and you would not notice until it reached a client. A model that says “I am not certain about this part” is a model you can actually delegate to.
You now control the effort
The second change is a new control in Claude and Cowork, available on every plan. Next to the model selector you can choose how much effort Claude puts into a response.
Turn it up and Claude thinks more, more often, for a better answer on a hard problem. Turn it down and it replies faster and uses your limits more slowly for the quick stuff. Opus 4.8 defaults to high, which Anthropic judges the best balance for most work, with “extra” and “max” settings above that for difficult or long-running tasks.
In practice this is a small habit, not a feature you need to study. Drafting a quick reply, summarising a document, tidying a list: leave it low and move on. Building a pricing model, working through a thorny contract, planning a launch: turn it up and let it take the time. The dial is there when you want it.
What to ignore
Here is the part the launch-day coverage tends to bury. The version number is not the thing standing between your business and useful Ai.
The numbers say so plainly. 64% of UK organisations now say they use Ai in some form, up from 52% a year ago. But only 24% have reached advanced use, where Ai is actually part of how the business runs and makes decisions, and that figure rose by a single percentage point last year. The models have moved fast. Adoption inside real businesses has barely moved at all.
That gap does not close because Opus went from 4.7 to 4.8. It closes when someone sits down with your actual work, the reporting, the admin, the email, the quoting, and builds the prompts and workflows that do that work the same way every time. The model is the engine. Most businesses are still missing the wiring.
So enjoy the upgrade. It is free, it is already in your account, and it is genuinely better at telling you when it is unsure. Then ask the more useful question: what in your week could this actually take off your plate, and who is going to build that with you.
Where Breezi fits
Membership gives you bespoke hours with a senior Ai team every month: you bring the problem, we build the prompt, the workflow, or the automation that handles it. Lahat Creative, a UK marketing agency, reclaimed up to 30% of their working hours every month doing exactly that. Consultancy does the same at the scale of a team, a department, or a whole organisation, the way one of our founders did inside a global financial platform. A new model is a good day. A model that quietly runs the parts of your business you never have time for is the point.
Free guide
Not sure where to start?
Get the cheat sheet.
ChatGPT, Claude, Gemini, Microsoft Copilot, Perplexity, and NotebookLM. What each one is for, how to set it up, and what to do in your first hour. Plain English. Ten pages. No fluff.
Sent
Your guide is downloading.
Didn’t start? Download it manually here.
If you opted in, the next Tuesday Memo lands in your inbox on Tuesday morning.
You’ll also get the occasional email about Ai that’s actually worth reading. Unsubscribe whenever.