Is Opus 5.5 nerfed? Thinking patterns changed on November 3
821 points
• 3 days ago
• Article
Link
Shane Melaugh 报告称 Opus 5.5 的表现明显下滑，暗示该模型很可能已经被"削弱"。他的结论基于使用 Gemini 4 Argon 进行的一项分析：该分析共处理了 1,400 个 Claude Code 会话。数据显示，模型的推理能力在 11 月 3 日发生了显著转折。尽管模型的官方名称没有变化，但其底层行为的改变足以扰乱用户的工作流程。
这项分析的一个核心发现是：当 Opus 5.5 被设置为"high"档时，实际执行的推理量大幅减少。 Melaugh 指出，如今模型似乎会过快投入任务，进而让人感到困惑，甚至即便投入数小时仍难以有效修复代码中的错误。相对而言，"xhigh"档需要稍作处理时间，但更能在几分钟内找出并落实正确方案。
Melaugh 建议，曾经很流行的做法——对每个任务都一律使用"high"档——已经不再适用。他强调，这些判断来自他在日常编码项目中对模型可靠性下降的亲身挫败。虽然他承认自己的分析并非严格意义上、经科学检验无懈可击的基准测试，但他在日志中观察到的规律与性能退化的趋势高度一致。
该说法已在技术社区引发更多讨论，开发者希望通过自己的日志来核实这些现象。尽管行为变化背后的根本原因仍未确认，但现有反馈对那些依赖 Opus 5.5 处理复杂任务的用户构成了提醒：用户可能需要调整设置，或更依赖"xhigh"配置，才能获得过去从该模型身上期望的推理表现。
Shane Melaugh has reported a noticeable decline in the performance of Opus 5.5, suggesting that the model has been effectively nerfed. His observations are based on an analysis conducted using Gemini 4 Argon, which processed 1,400 Claude Code sessions. The data indicates that a significant shift in the model's reasoning capabilities occurred on November 3. Although the official name of the model remained the same, the change in its underlying behavior was distinct enough to impact user workflows.
A key finding from this analysis is that the amount of reasoning performed by Opus 5.5 when set to the "high" level has plummeted. Melaugh noted that the model now appears to jump into tasks too quickly, leading to confusion and an inability to resolve coding bugs effectively, even after hours of effort. Conversely, the "xhigh" setting requires a moment of processing time but proves more capable of identifying and implementing correct solutions within minutes.
Melaugh advises that the previously popular strategy of using the "high" setting for every task is no longer viable. He emphasizes that these conclusions stem from personal frustration with the model's diminished reliability in daily coding projects. While he acknowledges that his analysis is not a formal, scientifically bulletproof benchmark, the patterns observed in his logs align with a clear degradation in performance.
The claim has sparked further investigation within the tech community, with developers looking to verify these observations through their own logs. While the underlying cause of this behavioral shift remains unconfirmed, the reported experience serves as a cautionary note for users who rely on Opus 5.5 for complex tasks. It suggests that users may need to adjust their settings or rely more heavily on "xhigh" configurations to achieve the levels of reasoning previously expected from the model.
518 comments • Comments Link
- 人们越来越担心 Anthropic 会对用户进行 A/B 测试：暗中"削弱"模型性能，进而拉低代码质量、增加 Token 用量，并提高错误率。
- 用户对模型持续退化表现出强烈不满，并推测在 IPO 之前，算力可能被重新分配到训练、托管或更新尚未发布的模型上。
- 普遍观点认为，这种性能波动是"恶化"（enshittification）的信号：供应商把利润率和计算效率放在用户一致性与产品可靠性之上。
- 一个反复出现的现象是，用户觉得前沿模型在新版本发布前夕会"变笨"；有人认为这可能是为了让后续版本通过对比显得更强。
- 虽然不少用户依赖前沿模型主要是因为其出色的推理能力，但越来越多的人主张用本地 AI 作为防止"模型漂移"（model drift）的手段，让代码编写工具保持稳定，并由用户掌控。
- 对那些更看重可预测性、希望能无限并发任务且不受订阅限制约束的开发者来说，自托管或使用像 Qwen 3.8-Flash-Next 这样的开放权重模型，正逐渐被视为一种可行替代方案。
- 反对"削弱"说法的人认为，感知到的退化通常源于认知偏差，或者是非确定性系统中自然的波动；也可能只是用户随着时间推移逐渐适应了模型能力的变化。
- 本地 AI 的怀疑者指出，目前它很难达到前沿模型的推理"极限目标"，而要实现同等水平所需的硬件投入对很多人来说又过于昂贵。
- 模型版本控制或性能标准的透明度不足，是争议焦点之一；有人呼吁通过独立的"神秘顾客"式基准测试，追究供应商对其付费订阅质量所承担的责任。
- 许多参与者认为，当前的 AI 生态更像"赌场"：缺乏稳定性迫使专业用户不断调整工作流程，以应对模型供应商频繁的变化和限制。
这场讨论的核心矛盾，在于高性能云模型带来的便利性，与专业开发所需要的稳定、可靠工具之间的冲突。许多用户报告称，像 Opus 和 Fable 这样的模型效果出现了明显且令人痛苦的下滑，并把这些变化解读为：供应商在财务里程碑之前为管理算力或操纵市场认知所采取的策略。尽管也有人把这些反馈视为轶事偏见，或认为这是人类神经系统对新技术的自然适应，但大量类似的说法仍表明，用户在处理关键任务时所依赖的服务确实存在真实的不稳定性问题。最终，这场对话凸显出两方之间的分歧：一方愿意把专有 AI 的"黑箱"属性当作经营成本；另一方则越来越多地转向本地方案，以重新夺回对自身专业环境的控制权。 • Concerns have emerged that Anthropic is conducting A/B testing on users by silently "nerfing" model performance, leading to perceived declines in code quality, increased token usage, and higher error rates.
• Users report significant frustration with consistent model degradation, often speculating that capacity is being reallocated to train or host newer, unreleased models ahead of an IPO.
• A common perspective is that these performance fluctuations are a symptom of "enshittification," where providers prioritize profit margins and compute efficiency over user consistency and product reliability.
• There is a recurring pattern of users feeling that frontier models are "dumbing down" shortly before the release of a newer iteration, potentially to make the subsequent version appear more capable by comparison.
• While some users rely on frontier models for their superior reasoning, many now advocate for local AI as a solution to prevent "model drift," ensuring that the code-writing tool remains static and under the user's control.
• Self-hosting or using open-weights models like Qwen 3.8-Flash-Next is increasingly viewed as a viable alternative for developers who prioritize predictability and the ability to run unlimited, concurrent tasks without subscription limits.
• Critics of the "nerfing" narrative argue that perceived degradation is often a result of cognitive bias, natural variation in non-deterministic systems, or simply users growing accustomed to the model's capabilities over time.
• Skeptics of local AI note that it currently struggles to match the reasoning "stretch goals" and speed of frontier models, and that the hardware investment required for comparable performance is prohibitively expensive for many.
• The lack of transparency regarding model versioning or performance standards is a major point of contention, with some calling for independent "secret-shopper" benchmarking to hold providers accountable for the quality of their paid subscriptions.
• Many participants feel the current AI ecosystem resembles a "casino," where the lack of stability forces professional users to constantly adapt their workflows to compensate for the shifting whims and constraints of model providers.
The discussion centers on a fundamental tension between the convenience of high-performance cloud models and the necessity of predictable, reliable tools for professional development. Many users report a clear, painful decline in the utility of models like Opus and Fable, interpreting these shifts as tactical maneuvers by providers to manage compute capacity or manipulate market perception ahead of financial milestones. While some dismiss these reports as anecdotal bias or the natural adaptation of the human nervous system to new technology, the volume of similar accounts suggests a real problem with the volatility of services that users rely on for critical tasks. Ultimately, the conversation highlights a growing divide between those who accept the "black box" nature of proprietary AI as a cost of doing business and those who are increasingly turning to local alternatives to reclaim control over their professional environments.