Sonnet: definition + examples

Sonnet is a model family within Anthropic's Claude lineup, positioned as a middle tier between the smaller, faster Haiku and the larger, more capable Opus. As of 2026, the current generation is Claude 3.5 Sonnet, released in mid-2025 as an incremental improvement over Claude 3 Sonnet (March 2024). Architecturally, Sonnet models are decoder-only transformers with approximately 70 billion parameters, using grouped-query attention (GQA) with 32 key-value heads and 64 query heads, a context window of 200,000 tokens, and a vocabulary size of 100,000 tokens trained via BPE tokenization. They are trained on a mixture of licensed web data, books, scientific papers, and code, with a cutoff date around early 2025. The training process uses a combination of next-token prediction on ~10 trillion tokens, followed by RLHF with constitutional AI (CAI) to align outputs with helpfulness, honesty, and harmlessness. Sonnet is distinguished by its ~3x higher inference throughput and ~2x lower cost per token compared to Opus, while achieving comparable performance on benchmarks like MMLU (88.7%), HumanEval (84.3%), and GSM8K (92.1%). It supports system prompts, tool use (function calling), structured output (JSON mode), and multimodal input (images, PDFs, tables). Common use cases include customer support chatbots, code generation and review, document summarization, data extraction, and RAG pipelines. Compared to alternatives, Sonnet offers a better accuracy-speed trade-off than GPT-4o (which is faster but slightly less accurate on reasoning tasks), and a more cost-effective alternative to Gemini 1.5 Pro for long-context tasks. A common pitfall is assuming Sonnet's performance on benchmarks translates directly to specialized domains (e.g., legal or medical reasoning), where it may still require fine-tuning or retrieval augmentation. Another pitfall is underestimating latency for real-time applications: while fast, Sonnet's first-token latency (~300ms) can be too high for voice-based interfaces, where Haiku is preferred. As of 2026, Claude 3.5 Sonnet is widely deployed via Anthropic's API and Amazon Bedrock, with a reported 200,000+ active developers and pricing at $3 per million input tokens and $15 per million output tokens. It remains Anthropic's flagship model for production use, with ongoing research into longer context windows (targeting 1M tokens) and improved tool-use reliability.