Claude Sonnet 4.6 — Stronger Coding, Reasoning & Extended Context
Anthropic has launched Claude Sonnet 4.6, an updated mid-tier model that improves on its predecessor across coding, multi-step reasoning, and extended-context tasks. The model is available immediately via the API as claude-sonnet-4-6, in Claude.ai (Pro and above), and through Amazon Bedrock and Google Vertex AI. Pricing is unchanged from Sonnet 4.5: $3 per million input tokens and $15 per million output tokens.
Benchmark improvements over Sonnet 4.5
- SWE-bench Verified — up from 62.8% to 68.4%, approaching Opus 4.6 territory on software engineering tasks
- GPQA Diamond — improved reasoning on graduate-level science questions
- HumanEval and LiveCodeBench — gains in both pass@1 and instruction-following for code generation
- 200K token context window — maintained from Sonnet 4.5, with improved faithfulness on long-context retrieval tasks
Anthropic describes Sonnet 4.6 as the recommended default for most production use cases — offering a balance of capability and cost that positions it above Haiku 4.5 for complex agentic workflows without the compute cost of Opus 4.6. Claude Code CLI and the VS Code extension have both been updated to use Sonnet 4.6 as the default model.