The article's central thesis is that Opus 5 isn't a bad model, but it was tuned for a slightly different job than what power users actually give it. Concrete regressions include premature tool calls, poor recovery from mistakes, declaring victory on partial refactors, and over-explaining trivial edits while under-explaining risky ones.
Argues the honest answer is 'worse at some things, better at others' — the more interesting question is why the gap between eval scores and lived experience keeps widening with each frontier release. Points to evals being frozen artifacts while real work is not.
Commenters cite specific numbers: repos where Opus 4.1 finished a migration in one shot and Opus 5 needs three, and tool-call malformed-JSON rates rising from ~0.4% to ~1.1% in agent frameworks. Long refactors that used to complete autonomously now ask for confirmation on steps the model previously just did.
Staff acknowledged some of the reported regressions in-thread and pointed users to the model card's known-regressions section. The editorial frames this document as 'quietly one of the more honest documents any frontier lab has published this year.'
Notes that Opus 5 posts Anthropic's best-ever scores on SWE-bench Verified, Terminal-Bench, and internal agentic harnesses, yet paying users still report worse workflows. The structural reason offered: evals are frozen, cleanly-specified artifacts, whereas real coding work is messy and open-ended — so improvements on the benchmark set don't transfer.
A post titled *Why does Opus 5 feel worse to work with?* hit the front page of Hacker News with 880 points, and the comment section read like group therapy for anyone who's shipped an agent in the last six months. The author walks through concrete regressions in long-horizon coding tasks: more premature tool calls, worse recovery from its own mistakes, a tendency to declare victory on partially-completed refactors, and — the one that stung most commenters — a subtle drift toward over-explaining trivial edits while under-explaining the risky ones.
None of this shows up on the headline evals. Opus 5 posts the best scores Anthropic has ever shipped on SWE-bench Verified, Terminal-Bench, and the internal agentic harness suites — and still, a large and vocal chunk of paying users say their workflows got worse the day they upgraded. The post's central claim isn't that Opus 5 is a bad model. It's that the model was optimized for a slightly different job than the one experienced users actually give it.
The HN thread piled on with specifics: repos where Opus 4.1 finished a migration in one shot and Opus 5 needs three; agent frameworks where the tool-call malformed-JSON rate ticked up from ~0.4% to ~1.1%; long refactors where the model now asks for confirmation on steps it used to just do. Anthropic staff replied acknowledging some of the reports and pointing at the model card's known-regressions section, which is quietly one of the more honest documents any frontier lab has published this year.
The interesting question isn't whether Opus 5 is worse — the honest answer is *worse at some things, better at others*. The interesting question is why the gap between eval scores and lived experience keeps widening with each release, and why nobody seems able to close it.
Three structural reasons show up over and over in the comments. First: evals are frozen artifacts, and real work isn't. SWE-bench tasks have a clean problem statement, a fixed repo state, and a deterministic grader. Your Tuesday-afternoon refactor has none of those. The moment a model gets tuned to squeeze another two points out of a static benchmark, it can quietly lose ground on the noisier distribution you actually live in.
Second: safety and helpfulness post-training pull in opposite directions on agentic tasks. A model that hedges more, asks permission more often, and refuses ambiguous instructions more aggressively will score better on trust-and-safety evals and worse on "just finish the damn PR" tasks. Every lab is doing this rebalancing every few months, and every time they do, someone's carefully-tuned agent loop breaks in a way that's hard to A/B because the failure mode is subjective.
Third — and this is the one the post argues most forcefully — the harness assumptions bake in. When Opus 4 shipped, framework authors learned its quirks: how much context it wanted per tool call, how it handled partial JSON, when it needed a nudge vs. when it needed to be left alone. Opus 5 has slightly different quirks. If your harness still assumes the old ones, you're not really running Opus 5 — you're running Opus 5 through an Opus 4 costume, and the seams show.
The community reaction split roughly three ways. The largest camp said "pin the old model until you have time to re-tune." A second camp said the regressions are real but temporary — the same thing happened at 3→3.5 and 3.5→4, and by the .1 release most complaints faded. A smaller camp argued this is the beginning of a real ceiling: that post-training now trades off enough capabilities against each other that "better on average" is starting to mean "worse on your specific workload" for a meaningful fraction of users.
If you ship anything that calls a frontier model in a loop, three things are worth doing this week. First, don't upgrade production agent loops on release day — pin the previous minor version for at least two weeks and let someone else discover the sharp edges. The cost of being on the very latest model is almost never worth the debugging tax when your users start filing tickets about "the AI got dumber."
Second, if you *have* upgraded and things feel off, look at your harness before you blame the model. Trim your system prompt — the older, longer version was tuned for a model that wanted more scaffolding. Tighten your tool schemas; Opus 5 is more literal about required fields than 4.1 was, and vague descriptions get vague calls. Cut the context you're feeding in half and see if quality goes up (it often does). The post's most quoted line — "the model didn't get worse, your prompt got older" — is annoyingly close to the truth for most reported regressions.
Third, build the eval you actually care about. Every team running agents in production should have a private set of 20-50 real tasks from their own workload, graded by a human, that they can run against every new model checkpoint. Public benchmarks will keep going up. Your private eval is the only thing that tells you whether the number going up matches the experience getting better.
The "feels worse" complaint is going to keep showing up with every major release, from every lab, because the underlying tension is structural: labs optimize for the average of a very wide distribution, and power users live in the tails. The teams that stay productive across model transitions are the ones who treat the model as one component in a system they own — not as a magic box whose new version is always an upgrade. Opus 5 is probably the best coding model shipping today. It's also, for a non-trivial number of workloads, a downgrade until you retune. Both things can be true, and pretending otherwise is how you end up rewriting your agent framework at 2am.
Top 10 dev stories every morning at 8am UTC. AI-curated. Retro terminal HTML email.