Is Opus 5.5 nerfed? Thinking patterns changed on November 3

Users of Anthropic’s Claude Opus 5.5 model report a sudden drop in code quality, reliability, and “thinking” behavior, leading many to suspect the model has been quietly “nerfed,” especially ahead of new model launches and the company’s IPO. Others counter that benchmark trackers and API-based evals show no clear degradation, suggesting normal variance, harness changes, or user bias rather than a deliberate downgrade. The debate highlights growing distrust of opaque model updates, concerns about A/B testing and cost-cutting on paid tiers, and renewed interest in open-weight or self-hosted models that offer more predictable behavior even if they are less capable.

Perceived “nerfing” of Opus 5.5 and related models

  • Many users report a clear drop in Opus 5.5 quality starting around Nov 3–7: more basic mistakes, skipped steps, worse code, repeated tool-call failures, ignoring prior instructions, and “hungover” / lazy behavior.
  • Some also report Fable and other Anthropic models becoming slower, more verbose, less reliable, and hitting safety rails more often.
  • Several mention needing to raise effort level (e.g., to “high”/“x-high”) to get results that previously worked on “low”/“medium,” and burning through usage limits much faster.

Explanations proposed (speculative within the thread)

  • Capacity / cost: reallocate compute to train or launch new models; heavy quantization or distillation to cheaper variants; reduce thinking tokens or context; prioritize “fast” or enterprise/API traffic over subscription users.
  • Marketing / IPO: weaken old models before new releases so upgrades feel dramatic; planned obsolescence / “enshittification.”
  • A/B testing: silent experiments on harness prompts, safety layers, routing, and quant levels; some claim subscription users are more affected than API or cloud-hosted versions.
  • Bugs: past examples cited of harness or TPU-related regressions, cache handling issues, or prompt changes hurting performance.

Skepticism and counterpoints

  • Multiple commenters say their own long-running usage shows no significant degradation beyond normal variance.
  • Public trackers (e.g., Livenerf, nerf-bench, other benchmarks) and some private eval suites reportedly see no clear drop in Opus 5.5 performance, especially via API.
  • Some attribute the pattern to human psychology: novelty high when models launch, then expectations rise and flaws become more salient; confirmation bias around “nerf” narratives.
  • A few small evals show at most modest or statistically unclear shifts.

Broader themes: trust, determinism, and business risk

  • Strong frustration that paid tools change unpredictably, breaking workflows, security reviews, and long-running projects.
  • Local/open-weight models praised for stability and version control, despite lower raw capability and high hardware costs; others argue frontier cloud models remain far ahead in quality and overall productivity.
  • Calls for:
    • Transparent changelogs for models and harnesses.
    • Independent, continuous nerf-bench-style tracking (including subscription “secret shopper” testing).
    • Clear separation of new versions vs silent changes to existing SKUs.
  • Overall, whether a real, measurable nerf occurred remains unclear in the thread; what is clear is widespread loss of confidence in model stability.