Near-Opus performance at one-fifth the cost with 1M-token context
Product AnnouncementAnthropic released Claude Sonnet 4.6, delivering performance previously requiring Opus-class models at Sonnet pricing ($3/$15 per million tokens). Features 1M-token context (2x previous Sonnet), 72.5% on OSWorld (nearly 5x improvement in 16 months), and became the default model for Free and Pro users. Matched Opus 4.5-class performance across coding, agents, and professional work.
The key value proposition: capabilities that previously required Claude Opus 4.5 (at $5/$25) are now available at Sonnet 4.6 pricing ($3/$15). This 40% cost reduction for equivalent capability accelerates enterprise adoption, where per-token costs directly impact ROI calculations. The pattern mirrors previous generational shifts where Sonnet catches up to prior Opus.
Sonnet 4.6 achieves 72.5% on OSWorld, the real-world computer interaction benchmark. For context, Sonnet 3.5 scored ~15% in October 2024. This nearly fivefold improvement in 16 months demonstrates the rapid advancement in AI computer use capabilities. At 72.5%, Sonnet 4.6 approaches GPT-5.4's 75% (on OSWorld-Verified).
The 1M-token context window (in beta) is double the previous 500K for Sonnet, matching Opus 4.6 and Gemini 2.5. For agentic and enterprise use cases, this removes a key limitation that previously forced users to choose Opus for long-context work.