Everyone Says Opus 5 Got Worse. The Data Disagrees

TL;DR: On 14 August a post titled “Why does Opus 5 feel worse to work with?” reached 853 points on Hacker News. The complaint is specific and widely shared: it stopped asking questions and started making assumptions. But the post contains no measurements — its author labels their own explanation “baseless speculation” — and the one public benchmark that tests exactly this puts Opus 5 four times ahead of Opus 4.8. Before you downgrade anything, there’s a setting worth checking.

What people are actually complaining about

The post is short and lands three specific charges at Opus 5. It doesn’t stop and ask when your intent is unclear. It makes assumptions without verifying them. It changes plans without asking first. The author’s summary is that it needs “careful babysitting”, where Opus 4.7, 4.8 and Fable would ask for clarification instead.

Two things about that post matter as much as its contents. First, it resonated — 853 points is a very big Hacker News day, and the comments are full of people saying the same thing. That makes the perception real and worth taking seriously.

Second, it presents no data at all. The author is upfront: the claim is “in my opinion and that of the colleagues I’ve spoken with”, and the section proposing an explanation is titled, in the author’s own words, “Baseless speculation”. That doesn’t make them wrong. It does mean nobody should treat this as a measurement, and a lot of the coverage repeating it has quietly dropped that caveat.

Anthropic has not publicly responded. We checked its newsroom on 15 August 2026: nothing addressing model quality, and the most recent Opus 5 item is the launch announcement from 24 July.

What the data actually says about Opus 5

There is one public benchmark aimed squarely at this complaint. SlopCodeBench doesn’t test whether a model can solve a problem in one go — it tests whether a model can keep a codebase healthy as new requirements arrive in stages. Each challenge has multiple “checkpoints”, and the model doesn’t see the whole problem up front. To pass a checkpoint “strictly”, it has to make the new thing work and not break anything from earlier checkpoints.

That is, near enough, a machine version of “it makes assumptions and quietly wrecks things”.

Dex Horthy of humanlayer ran three Claude models through a 17-checkpoint subset in late July and published the results:

Model	Strict pass rate
Opus 5	24% (4 of 17)
Opus 4.8	6% (1 of 17)
Sonnet 5	6% (1 of 17)
Opus 4.6 (original paper)	17%
GPT-5.4 (original paper)	11%

Opus 5 came first, by four times over the model everyone says it’s worse than.

The verbosity complaint holds up better, but with a twist. Opus 5 wrote roughly five times as many functions as Opus 4.8 and produced 29,065 lines against about 9,000 — except 51% of that was test code, versus 11–24% for the others. And on the measure of pointless code, Opus 5 looks best: only 14.9% of its functions were used once and never again, against 49.1% for Opus 4.8 and 71.5% for Sonnet 5.

Read honestly, that’s a model writing more, much of it tests, with less throwaway padding — not a model that got dumber.

Two caveats we’d want stated if this were about us: it’s a small subset run by one person, published in late July rather than this week, and every model scored badly. Horthy’s own conclusion is the sober one — that “today’s models can’t be relied on to run lights-off without steering”, and he’d want 80%+ before trusting one unsupervised.

The bit nobody’s connecting

Here’s what makes us cautious about the “the model got worse” story. The loudest complaint — it doesn’t ask me any more — describes a change that demonstrably happened, but in the tool, not necessarily the model.

Claude Code’s own changelog shows its defaults moving steadily toward acting without checking in: subagent forking switched on by default, background agent spawns made the default, auto mode made available without opting in, and the permission setting formerly called `”default”` renamed to `”manual”` — a small word with a lot in it. We wrote about that shift on 10 August, before this row started: Claude Code is about to stop asking you.

To be clear, this is our inference and not an established fact — we haven’t seen anyone demonstrate that the two are connected, and it’s entirely possible Opus 5’s manners changed too. But if your AI stopped asking permission around the time the defaults changed to stop asking permission, that’s worth ruling out before you conclude the model is broken.

Why it matters if you don’t code

Non-developers are the people most exposed to this kind of row, for a simple reason: you’re told to always pick the newest and best model, you have no way to check the claim, and when the internet says the new one is worse, you have no way to check that either.

This one is a useful case study, because both sides are half right. The behaviour people are describing is real. The conclusion they’re drawing from it — downgrade to the old model — is the one thing the available data doesn’t support.

It also lands next to an argument we’ve been watching. Lovable published an essay on 11 August arguing that the model picker is a dead end and non-developers shouldn’t be choosing models at all. Whatever you make of that coming from a vendor, days like this are its best evidence.

What to do

1. Check your permission mode before you change anything else. In Claude Code, look at whether you’re in auto mode or manual. If it stopped asking you, this is the first thing to rule out.
2. Don’t downgrade on vibes. If you were about to switch back to Opus 4.8 because of a viral post, note that the only public measurement of that exact complaint has 4.8 at 6% and Opus 5 at 24%.
3. Ask for a plan before it builds. “Tell me what you’re going to do before you do it” restores the checkpoint people are missing, whatever the cause.
4. Judge on your own work, not a benchmark or a thread. Run the same real task on both and keep the one that does it better. Yours is the only score that matters.
5. If you’re on an app builder, this mostly isn’t your problem. Lovable, Bolt and Horizons pick the model for you.

Who should care (and who shouldn’t)

• Using Claude Code directly: the most affected group — check your permission mode, then judge for yourself.
• Building a revenue app and considering a downgrade: hold off. Test both on your own work first.
• Using Lovable, Bolt, Replit or Horizons: essentially nothing to do; the model is chosen for you.
• Learning, or not started yet: ignore this entirely. Model rows are a spectator sport at this stage — the quiz will get you further in 60 seconds.

Our take

We think the honest reading is that people are describing something real and diagnosing it wrongly. “It stopped asking me” is a behaviour change you can point at in a changelog; “the model got worse” is a much bigger claim, and the one public number aimed at it goes the other way.

What we’d genuinely take from this week is duller and more useful than either camp: every model tested scored badly at keeping a codebase healthy over successive changes — the best of them managed 24%. That’s the real finding. Whichever model you use, it needs steering, and the people quietly succeeding with these tools are the ones who check the work rather than the ones who picked the right version number.

Not sure which AI tool fits what you’re building? Take the 60-second Vibe Coding Tool Finder quiz →

Some links on this page are affiliate links. If you buy through them we may earn a commission at no extra cost to you.

FAQ