Anthropic has released two new large language models, Claude Opus 4 and Claude Sonnet 4, which introduce "hybrid reasoning" capabilities to merge quick responses with extended step-by-step thinking. These models succeed the Claude 3.x generation and retain a 200,000-token context window while offering improved safety and agentic abilities.
- On the SWE-bench Verified benchmark, Opus 4 achieved 72.5% accuracy and Sonnet 4 reached 72.7%, outperforming OpenAI's GPT-4.1 (54.6%) and Google's Gemini 2.5 Pro (63.2%).
- Anthropic reports that Sonnet 4 exhibits up to eight times lower reward hacking compared to the previous Sonnet 3.7 model.
- Independent benchmarks show Opus 4 placing third in Aider tests with 72% accuracy and Lovable reporting a 40% speed increase for Sonnet 4.
- The models are designed to handle long-running tasks, with Opus 4 capable of autonomous coding sessions lasting several hours without losing context.
The release aims to deliver significant gains in coding, reasoning, and long-form content handling while maintaining the conversational safety standards associated with the Claude brand.