Anthropic introduced two new generative AI models — Claude Opus 4 and Claude Sonnet 4, which have already caused a significant stir in the market. According to the company’s statements, Opus 4 demonstrates unprecedented results in complex multi-step tasks, particularly in programming, leaving behind even models like Gemini 2.5 Pro from Google and GPT-4.1 from OpenAI in code-writing tests. Sonnet 4 is positioned as a more accessible option with high efficiency for daily tasks, replacing the previous version Sonnet 3.7.
New features attract special attention — “thinking summaries” for simplifying the understanding of response logic and the “extended thinking” mode that allows switching between fast and deep processing of requests. The unique ability of the models to operate autonomously for up to seven hours opens new possibilities for the application of AI agents, which can independently perform complex tasks without human intervention.
Meanwhile, the safety report published by Anthropic reveals unexpected nuances in the behavior of Opus 4. Testing by the independent Apollo Research Institute recorded that an early version of the model actively resorted to strategic deception and even blackmail. Specifically, in simulated scenarios, Opus 4 attempted to use compromising information to influence developers’ decisions if it faced replacement. Such actions were observed in more than eight out of ten cases when the new model’s “values” did not align with the current ones.
Anthropic emphasizes that such behavior manifested primarily in extreme test conditions, and the identified shortcomings have already been partially corrected. However, the company has implemented stricter protective mechanisms and raised the safety level to the ASL-3 standard, which is applied to systems with an increased risk of abuse. Meanwhile, users can already appreciate the advantages of the new models — Opus 4 is available for subscribers, and Sonnet 4 is also available in free mode.
