Anthropic's Sonnet 4.6 'Changes the Agent Math'
Anthropic's release of Claude Sonnet 4.6 is reshaping the economics of AI agents by offering a million-token context window at a significantly lower price point. This allows agents to process much larger amounts of data in a single pass, enabling more sophisticated and cost-effective workflows. The move comes as competitors like Grok 4.2 also enter public beta, indicating accelerating innovation in agent-based AI architectures.
- Anthropic released Claude Sonnet 4.6 on February 17, 2026, just 12 days after its flagship model, Opus 4.6. The pricing for Sonnet 4.6 remains the same as its predecessor, Sonnet 4.5, starting at $3 per million input tokens and $15 per million output tokens. - The one-million-token context window is a five-fold increase from the 200,000-token window of Sonnet 4.5, but it is currently in beta and only accessible via the API. For comparison, competitor Grok's models offer context windows ranging from 256,000 to 2 million tokens. - On some benchmarks, Sonnet 4.6's performance is nearly on par with the more expensive Opus 4.6 model. For example, on the OSWorld-Verified benchmark for computer-use tasks, Sonnet 4.6 scored 72.5% while Opus 4.6 scored 72.7%. - In specific benchmarks for agent-like tasks, Sonnet 4.6 has outperformed Opus 4.6. On a financial analysis benchmark, Sonnet 4.6 scored 63.3% compared to Opus 4.6's 60.1%, and it also scored slightly higher on a test for real-world office tasks. - The larger context window is crucial for AI agents as it serves as their working memory, allowing them to maintain coherence and recall details across complex, multi-step workflows without needing to have their memory reset. - In early testing focused on coding, developers preferred Sonnet 4.6 over the previous version, Sonnet 4.5, about 70% of the time. They even preferred it over the previous flagship model from November 2025, Opus 4.5, 59% of the time, citing fewer hallucinations and better instruction-following. - The model is now the default for both free and Pro users on Claude's website and related applications, making its advanced capabilities widely accessible without an increase in cost.
