Anthropic introduced a new hybrid reasoning model on Thursday, Claude Opus 4.7, which reportedly reduces hallucinations and sycophancy compared to previous and other frontier AI models. This latest offering from Anthropic maintains its reputation as a safety-first AI company. The Opus 4.7 system card indicates less susceptibility to inaccuracies and sycophantic conduct compared to its predecessors, like Opus 4.6 and Sonnet 4.6, and even competing models like Gemini 3.1 Pro and Grok 4.20.
Claude Opus 4.7 shows substantial improvements in honesty, marked by reductions in significant omissions and moderate enhancements in factual accuracy and hallucination rates. Notably, its MASK honesty rate is 91.7%, slightly lower than the 95.4% Claude Mythos, yet better on other hallucination scores. Opus 4.7’s false premises honesty rate stands at 77.2%, outperforming all other Anthropic models except Claude Mythos at 80%.
In terms of sycophancy, Opus 4.7 maintains similar scores to previous Anthropic models, reflecting better behavior over rivals like Gemini 3.1 Pro and Grok 4.20. The model was audited using Anthropic’s Petri 2.0, a tool scoring bad behaviors on a 1-10 scale, where lower scores indicate better results.
Timothy Beck Werth, a tech editor with over 15 years of experience and specific expertise in consumer technology, contributed to this analysis.