Claude Opus 5: Why Users Say Anthropic's New Model Is a Downgrade
Claude Opus 5 posts strong benchmarks but many users report verbosity, overreach, and worse coding judgment than Opus 4.8. Here's what's going on.
What is Claude Opus 5, and why is it controversial?
Claude Opus 5 is Anthropic’s latest flagship model, released to strong scores across terminal coding, knowledge work, novel problem solving, agentic search, and computer use benchmarks, putting it roughly on par with GPT-5.6 on paper while costing noticeably less to run. The controversy comes from the gap between those benchmarks and actual usage: a large number of developers and knowledge workers report that Opus 5 feels worse than its predecessor, Opus 4.8, in daily work, especially coding.
TL;DR
- Benchmark parity, real-world complaints: Opus 5 matches or nearly matches GPT-5.6 on standard evaluations for coding, agentic search, and computer use, but many users say it performs worse than Opus 4.8 in practice.
- Verbosity is the top complaint: Users describe Opus 5 as scattered, hard to read, and exhausting to work with because it buries answers in excessive explanation instead of directly addressing the question asked.
- Overreach on coding tasks: Developers, including Theo (of t3.chat), report that Opus 5 treats minor issues as critical problems requiring massive, unnecessary code changes.
- One standout strength: Opus 5 appears unusually good at generating visually impressive game graphics in Three.js, producing environments that look close to AAA quality from a single detailed prompt.
- Gameplay lags behind graphics: Side-by-side tests show Opus 5 built worlds that look better than GPT-5.6’s output, but the actual game mechanics (controls, combat, collision) were often broken or backwards.
- Cost efficiency is real: Opus 5 hits rate limits slower and costs less than GPT-5.6 for similar benchmark performance, which matters for teams running high-volume agentic workloads.
- The reception looks like a rare regression: Multiple independent users, not just one disgruntled group, describe Opus 5 as feeling nerfed compared to Opus 4.8, which is unusual for a model marketed as a straightforward upgrade.
Other agents ship a demo. Remy ships an app.
Real backend. Real database. Real auth. Real plumbing. Remy has it all.
Why do the benchmarks and user experience disagree?
Benchmarks measure narrow, well-defined tasks under conditions that don’t always match how people actually use a model. Opus 5 scores well on structured evaluations like terminal coding challenges or agentic search tasks, where success is measurable and the task is bounded. Day-to-day use is messier. It involves back-and-forth conversation, ambiguous requests, and judgment calls about how much effort a problem actually warrants.
That’s where the complaints concentrate. Users don’t describe Opus 5 as incapable. They describe it as miscalibrated. It tends to escalate small issues into large ones, respond with more text than the question requires, and drift from the specific thing being asked. One widely echoed comment: Opus 5 “never answers the actual question you ask it.” Another compared its output style to having ADHD, saying it takes real effort just to read through a response and extract the useful part.
This kind of gap, strong on paper and frustrating in practice, suggests Anthropic may have tuned Opus 5 toward benchmark-style tasks or optimized for handling more edge cases automatically, at the cost of the concise, direct interaction style that made earlier versions pleasant to use for iterative work.
Is Opus 5 actually worse than Opus 4.8 for coding?
Multiple developers say yes, at least for their workflows. Theo, a well known figure in the coding tool space, described Opus 5 as the most annoying model he’s used, saying it treats every minor issue like a high severity problem requiring thousands of lines of code to resolve. Others echoed this, calling it “clearly a downgrade” and “nerfed, not upgraded” relative to 4.8.
This matters because coding is one of the areas where verbosity and overreach cause real friction. A model that rewrites large chunks of a codebase to fix a small bug creates more review work than it saves. A model that answers a narrow question with a sprawling explanation slows down anyone trying to move fast through an agentic coding session. Several users reported reverting to Opus 4.8 specifically because it “feels way better” for these workflows, which is a notable reaction for a model release meant to be an improvement.
It’s worth noting this isn’t universal. Not every user reported problems, and some likely had strong results, particularly for tasks that reward Opus 5’s apparent strengths. But the volume and consistency of the coding complaints, from independent, credible sources, stands out.
What is Opus 5 actually good at?
The one area where Opus 5 draws consistent praise is generating visually rich game environments, specifically using Three.js (3JS), a JavaScript library for 3D graphics in the browser. Several independent tests showed Opus 5 producing game worlds, from a Call of Duty style shooter to a fantasy environment styled after Elden Ring, with landscapes, lighting, and textures that looked close to AAA game quality from a single detailed prompt run for hours.
A side-by-side comparison using an identical prompt on GPT-5.6 Sol Ultra told a more nuanced story. GPT-5.6’s version had noticeably less impressive graphics but more functional gameplay: correct left/right controls, working combat, and more coherent character design. Opus 5’s version had backwards controls, unclear hit detection, and characters that looked more like rough shapes than finished models, but the environment itself, mountains, trees, water, looked dramatically better.
That split, striking visuals but shaky mechanics, matches the broader pattern in user feedback: Opus 5 seems to excel at generating impressive, high effort creative output in a narrow domain while struggling with the practical judgment needed for iterative, functional work.
Is Opus 5 worth using over Opus 4.8 or GPT-5.6?
For most day-to-day coding and knowledge work, the reporting suggests Opus 4.8 or GPT-5.6 remain the more reliable daily drivers. Multiple experienced users, across different use cases, independently gravitated back toward older models after trying Opus 5, which is a meaningful signal even accounting for the usual noise of early-release reactions.
Where Opus 5 does make sense is cost sensitive, high volume agentic workloads where its efficiency (similar benchmark performance to GPT-5.6 at meaningfully lower cost, with slower rate limit consumption) outweighs the friction of verbose or over-engineered responses. It may also be the better choice specifically for generative visual work like game prototyping, where its apparent strength in producing detailed 3D environments is a genuine differentiator, even if the resulting gameplay needs cleanup.
The honest takeaway is that “better” depends heavily on the task. A model that’s excellent at one thing and disappointing at most others isn’t a failed release, but it is a different kind of tool than a straightforward upgrade, and treating it as a drop-in replacement for Opus 4.8 across all workflows appears to be where most of the frustration comes from.
Frequently Asked Questions
Why do people say Claude Opus 5 is a downgrade from Opus 4.8?
Users report that Opus 5 is more verbose, harder to read, and prone to treating small issues as major problems requiring excessive code changes. Several developers and knowledge workers say they’ve gone back to Opus 4.8 because it produces more direct, usable results.
Does Opus 5 perform well on benchmarks?
Yes. Anthropic’s benchmarks show Opus 5 performing close to GPT-5.6 across terminal coding, agentic search, computer use, and novel problem solving, generally at a lower cost per use.
What is Opus 5 particularly good at?
Generating visually detailed game environments in Three.js. Multiple independent tests show it producing game worlds with graphics quality that approaches AAA titles from a single prompt, though the resulting gameplay mechanics are often rough or broken.
Is Opus 5 cheaper to use than GPT-5.6?
Based on available reporting, yes. Opus 5 offers benchmark performance close to GPT-5.6 while consuming rate limits more slowly and costing less, which matters for high volume or agentic use cases.
Should developers switch to Opus 5 for coding work?
Based on widespread user feedback, many developers are sticking with Opus 4.8 or other models for coding, citing Opus 5’s tendency toward overreach and unnecessarily large changes for small problems. Results may vary by workflow and task type.
