July 27, 2026
Slop Stars, Coding Scars
Benchmarking Opus 5 on SlopCodeBench
New test says Opus 5 still needs babysitting — and commenters are not impressed
TLDR: A new test found Opus 5 only modestly better at handling long, messy coding work, and none of the tested AI models finished cleanly. Commenters turned that into a roast, with some praising the benchmark itself but many dunking on Opus 5 as overhyped and sloppy.
A new coding test called SlopCodeBench was supposed to answer a very simple question: can today’s AI coding tools keep a software project healthy as new demands pile up over time? The answer, according to this run, is basically not really. Opus 5 technically won, scoring 24% on a small set of tasks, but the real crowd reaction was less “victory lap” and more "that’s it?" One commenter flat-out said Opus 5 is “not much of an improvement,” while another went nuclear and branded it an “overconfident stupid model” that produces too much messy code.
That tone pretty much defined the thread. The benchmark itself is getting praise because it tests something people actually care about: whether an AI can keep fixing and updating a project without breaking old stuff. One person even cheered, “finally the benchmark for me,” which is about as close to a standing ovation as benchmark nerds get. But the celebration quickly turned into model tribalism, nostalgia, and a little chaos. Fans of older versions were suddenly sounding like music snobs talking about a band’s early albums, reminiscing about the days when Opus 4 and 4.6 still had that magical “wow factor.”
And then came the mini food fight: why not test rival models too? One commenter volunteered to run GPT 5.6, GLM 5.1, or Kimi K3, while another challenged whether the whole write-up was even useful. So yes, Opus 5 “won” — but in the comments, the bigger story was disappointment, skepticism, and people arguing over which AI is the least sloppy intern.
Key Points
- •The article presents SlopCodeBench as a March 2026 long-horizon coding benchmark that reveals requirements over multiple checkpoints instead of all at once.
- •The author ran Opus 4.8, Sonnet 5, and Opus 5 on a 17-checkpoint subset of SlopCodeBench using the Claude Code harness.
- •Opus 5 achieved the highest reported score in the author's test subset at 24%, compared with the 17% strict pass rate cited for Opus 4.6 in the original paper.
- •The subset included three tasks: circuit_eval (easy, 8 checkpoints), database_migration (medium, 5 checkpoints), and dynamic_config_service_api (hard, 4 checkpoints).
- •Across all nine runs, none of the tested models completed any challenge with all checkpoints passing under the strict pass criterion.