On July 24, Anthropic released Claude Opus 5, its fourth new model in under sixty days. Sonnet 5, Fable 5, and Mythos 5 all landed in June. Four flagship-tier or near-flagship models, each targeting a different price point and capability band, shipped before most users had finished forming an opinion about the last one.
The reception was swift and sour. A blog post titled “Why does Opus 5 feel worse to work with?” hit the front page of Hacker News this week, racking up hundreds of comments from developers and knowledge workers who insist the new model is a regression. It’s verbose, they say. It misunderstands instructions. It produces output that requires more editing, not less. The benchmarks say it’s roughly on par with GPT-5.6. The users say it feels nerfed.
I don’t think Opus 5 is worse. I think the people using it are tired in a way that no benchmark can measure, and Anthropic — along with every other AI lab shipping on a monthly cadence — has structured its business to guarantee that exhaustion.
The Upgrade Treadmill Is the Product
Consider what a professional user of these models has been through since June. In roughly eight weeks, they have been asked to recalibrate their intuition about four different systems. Each one has different strengths, different failure modes, different verbosity levels, different quirks around when it hallucinates and when it refuses. A developer who built a workflow around Sonnet 5’s particular style of code generation in early June was, by late July, being nudged — or auto-upgraded — to something that behaves differently in subtle, maddening ways.
This is not a quality problem. It is a pace problem. The model might be perfectly capable, but capability is meaningless if the user hasn’t had time to develop the tacit knowledge that makes a tool feel like an extension of their own thinking. That tacit knowledge — the instinct for when to trust the output and when to double-check, the rhythm of prompting and refining — takes weeks to build. Anthropic is now shipping faster than that learning curve.
One engineer I spoke with described it this way in a Slack DM: “It’s like someone swapped out my keyboard switches every three weeks and then published a blog post about how the actuation force is technically lower.” He’s not wrong. The spec sheet improves. The experience degrades.
The Effort Dial Is an Admission
Buried in the Opus 5 release is a feature that tells you everything: the effort parameter. It defaults to high on the API and in Claude Code. Lower settings trade reasoning depth for speed and cost. Anthropic’s own documentation notes that Opus 5 generates 26% fewer tokens than Opus 4.8 at maximum reasoning depth.
Read that again. The flagship model, at its highest setting, produces less than its predecessor. That is not a bug. It is a design choice — one that makes economic sense for Anthropic, which charges by the token, and one that makes the model feel clipped to anyone who grew accustomed to the verbosity of 4.8.
The effort dial is a confession that the default experience is a compromise. It is a knob that lets the customer dial up to what they thought they were paying for. And most users, especially those on the Claude Pro tier where Opus 5 is now the strongest option, don’t touch API parameters. They get the default. The default is tuned for Anthropic’s margins.
Nobody Asked for a Model Every Two Weeks
The AI industry has convinced itself that shipping velocity is a proxy for progress. It is not. It is a proxy for competitive anxiety. Every lab is terrified of being perceived as falling behind, so they ship whatever is ready — and sometimes what isn’t — to keep the newsletter cycle churning. The result is a user base that has been conscripted into a permanent beta program, paying $20 or $200 a month for the privilege of debugging someone else’s release cadence.
This is not a call for slower innovation. It is an observation that the current pace serves the companies, not the customers. A model released every six months that users can actually learn would be more valuable than four models in sixty days that each feel like a stranger. But “we took our time” doesn’t generate the same investor deck slide as “four launches in Q3.”
The Hacker News thread is full of people trying to diagnose Opus 5’s technical deficiencies. They are looking in the wrong place. The model is probably fine. The problem is that Anthropic has optimized for a metric — releases per quarter — that correlates inversely with the thing users actually want, which is a tool they can trust for longer than a billing cycle.
The Real Regression
If there is a regression here, it is not in the model weights. It is in the relationship between the people building AI and the people using it. Every new release that arrives before the last one has been absorbed is a small act of disrespect. It says: your workflow is not our problem. Your learning curve is not our metric. Keep paying, keep prompting, and we’ll keep shipping.
Opus 5 doesn’t feel worse because it is worse. It feels worse because the user on the other side of the screen is running a mental benchmark that no one at Anthropic thought to measure: how much of my attention am I spending just keeping up with you? That number is going up. And at some point, even the best model in the world isn’t worth the cognitive overhead of learning to use it.