Claude Sonnet 4.6

von Anthropic

AI Model

Niedrige Vertrauensstufe

Claude Sonnet 4.6 is a high-performance, cost-efficient large language model developed by Anthropic, released in February 2026. It delivers near-frontier capabilities in software engineering, agentic tasks, reasoning, multimodal understanding, and computer use, often approaching or matching the performance of Claude Opus 4.6 on key benchmarks. The model supports up to 1M tokens (native context window) and excels in cost-performance trade-offs, making it ideal for scalable deployments. Claude Sonnet 4.6 was trained on a proprietary mix of publicly available internet data (up to May 2025), private datasets, opted-in user data, and synthetic data, with deduplication, classification, and filtering applied. It underwent RLHF and AI feedback fine-tuning to align with Claude’s Constitution (helpful, honest, harmless). The model is multilingual (with a ~4.4% average gap to English on GMMLU) and text-only. Safety evaluations show strong alignment, with a 99.38% harmless response rate on violative requests and low over-refusal rates (0.41%) on benign prompts. It was deployed under AI Safety Level 3 (ASL-3) protections, with no evidence of crossing CBRN-4 or AI R&D-4 thresholds. Notably, it outperforms Opus 4.6 in some areas (e.g., WebArena-Verified, CharXiv Reasoning, and MedCalc-Bench).