Is Opus 5.5 Better At Writing?

Opus 5.5 aims for more natural communication and improved adherence to user-provided writing rules, addressing common feedback about "slop." The key claims are that it writes more naturally and follows instructions better, allowing for personalized writing styles. A comparison between Opus 5 and 5.5 was conducted using a side-by-side review of their outputs. The initial task was to write a short FAQ article explaining AI evaluations (evals). Opus 5.5 was noted to have made fewer tool calls, though its impact on context reading was unclear. In the FAQ task, Opus 5.5 added "short" to the prompt, which was deemed unnecessary. The definition of an AI eval provided by Opus 5.5 described it as a process to measure an AI application, not a single test or score. A second writing task involved summarizing a blog post about "six things I tried with Jev," specifically for an audience skeptical of Jev due to the existence of classifiers. The prompt included instructions for a warm, professional tone and plain writing skills, with the model referencing a markdown file for guidance. The models were also anonymized to allow for later reveal. Opus 5's summary was described as too literal of the title. Opus 5.5's summary was initially critiqued for being too defensive in tone, assuming an objection from the reader. The writing was considered to be "AI writing," with complex sentences that required significant cognitive effort to understand. Opus 5.5's summary was later re-evaluated, with one participant finding its title better and the explanation of Jev's advantage (response time and no need for classifier building) clearer. However, it was still deemed to contain "slop" and went into unnecessary detail too early. Specific criticisms included overly specific numbers and made-up examples that didn't align with the original post. The confusion around what the "six tasks" actually were was a significant issue. The discussion then shifted to intentionally producing "slop" by attempting to craft a good Twitter reply. The context provided was a post about RLM being the most important function of a "pie harness," seeded with late interaction retrieval results, and involving setup and iteration loops. The prompt asked for a thoughtful reply from someone familiar with coding agents, inquiring about the author's conviction that this approach is superior to others without RLMs or similar harnesses. The reply was to be in the style of a LinkedIn influencer. Opus 5.5's attempt at a LinkedIn influencer style was described as "cringe." Both models produced output that was considered "slop," with one participant stating that "slop is not dead." The comparison of summaries for the Jev post highlighted that Opus 5.5 was generally better, but still not great. A further test involved generating a Twitter reply to a short post stating that something is good when it can be used to argue conflicting viewpoints. The prompt was to craft a reply that would be interesting and likely to engage the author, being short and thoughtful. Opus 5.5's response was not shorter than Opus 5's, and both were considered "really bad" and not detectable as human. The general consensus was that engagement bots are still a significant issue. The conversation returned to comparing Opus 5 and 5.5, with the realization that "slop" is still a problem, even if 5.5 might be slightly better in some cases. The comparison was framed as whether 5.5 produces "less slop." A final task involved generating a title and abstract for a conference proposal. The context provided was extensive, including a whole book. Opus 5.5's title was considered better, though still not ideal. Its abstract was deemed slightly better than Opus 5's, which was described as confusing and "classic slop." Opus 5.5's abstract was considered more parallel and better structured, but still "sucked." The similarity between the outputs of both versions was noted, raising questions about what changes were made to 5.5. Despite some perceived improvements in 5.5, the overall conclusion was that "slop is still here" and has not been solved. The review of six examples led to a feeling of pain, with the abstracts often being generic and lacking specific examples. The free resource on AI evals was then promoted.