Anthropic admits they nerfed their Claude model in August (twitter.com)

Anthropic acknowledged that in August it "nerfed" its Claude model — an intentional reduction in some capabilities or behaviors — and has now confirmed the change. While details on exactly which components were altered aren’t fully public, nerfs typically involve weight updates, stricter decoding constraints, additional safety filters, or reinforcement‑learning penalties that make the model less likely to produce risky or disallowed outputs. The company’s admission means that Claude’s behavior, benchmarks and user-facing performance can shift between releases even without a new major version number. This matters because capability regressions affect reproducibility, product integrations, and how the research community evaluates models over time. Developers and researchers relying on Claude for benchmarks, prompting strategies, or downstream apps may see differences in accuracy, hallucination rates, or jailbreak robustness tied to these safety-driven trade‑offs. Technically, nerfs can be implemented at multiple layers — fine‑tuning with constrained objectives, logits suppression, lower sampling temperatures, or external content filters — and each choice has different effects on fluency, factuality, and adversarial vulnerability. The episode highlights the need for transparent versioning and detailed model cards so users can understand safety-performance trade-offs and reproduce results reliably.