Claude 3.5 Sonnet outperforms prior flagship on benchmarks and agentic reasoning at mid-tier speed and cost

Anthropic released Claude 3.5 Sonnet on 21 June 2024, claiming frontier performance on standard benchmarks (GPQA, MMLU, HumanEval) while operating at 2x the inference speed of Claude 3 Opus and priced at the mid-tier level. In an internal agentic coding evaluation, it achieved 64% task completion versus Opus at 38%, demonstrating multi-step bug fixes and code generation. The model is distributed through multiple channels (web, API, AWS Bedrock, Google Vertex) and serves as the first release in a new Claude 3.5 family. For financial institutions, this signals continued generational improvement in inference economics and agentic capability, though the data is based on vendor benchmarks and internal evaluations without independent verification.