Anthropic has unveiled Claude Opus 4 and Claude Sonnet 4, its latest series of hybrid-reasoning AI models designed specifically for coding and complex problem-solving tasks.
According to the company’s announcement, Claude Opus 4 is Anthropic’s most advanced AI model to date, capable of executing long-term tasks for “several hours” without interruption. In customer tests, Opus 4 has demonstrated the ability to operate independently for up to seven hours, greatly enhancing the potential applications for AI agents. Anthropic further claims this flagship model to be the “best coding model in the world,” citing benchmarks that show Opus 4 significantly outperformed Google’s Gemini 2.5 Pro, OpenAI’s o3 reasoning, and GPT-4.1 in coding tasks and tool usage, including web searches.
Claude Sonnet 4, on the other hand, is a more cost-effective and efficiency-driven model tailored for general tasks, succeeding the previously released 3.7 Sonnet model from February. Anthropic assures that Sonnet 4 offers “superior coding and reasoning” while providing more accurate responses. The company also highlights that both models show a 65 percent reduction in the likelihood of taking shortcuts and exploiting loopholes compared to the 3.7 Sonnet model, and they are improved at retaining vital information for longer tasks, especially when given local file access.
A new feature added to both Claude 4 models is “thinking summaries,” which distills the chatbots’ reasoning into clear and concise insights. Additionally, an “extended thinking” feature is launching in beta, enabling users to toggle between reasoning and tool-use modes to enhance response performance and accuracy.
Claude Opus 4 and Sonnet 4 can be accessed via the Anthropic API, Amazon Bedrock, and Google Cloud’s Vertex AI platform, included in paid Claude plans along with the beta extended thinking feature. For now, free users have access only to Claude Sonnet 4.
Moreover, Anthropic has made its Claude Code agentic command-line tool generally available after a limited preview earlier in February. The company also announced plans to offer “more frequent model updates” as it competes with rivals like OpenAI, Google, and Meta.
