Claude Sonnet and Opus 4 (Executive Summary)

Anthropic released Claude Opus 4 and Sonnet 4 today, claiming the #1 spot for coding performance. There are going to be a lot of articles floating around with exaggerations and marketing talk, but here is an executive summary of everything you need to know.

Performance Numbers

Claude Opus 4:

• SWE-bench: 72.5% (world's best)
• Terminal-bench: 43.2%
• Sustained performance for hours on complex tasks
• $15/$75 per million tokens

Claude Sonnet 4:

• SWE-bench: 72.7% (matches Opus 4)
• 3x faster than Opus 4 for most tasks
• $3/$15 per million tokens

Two key slides from the announcement:

Key Technical Features

• Hybrid Architecture: Instant responses + extended thinking mode (up to 64K tokens)
• Extended Thinking with Tools: Can use web search, code execution during reasoning
• Parallel Tool Execution: Multiple tools simultaneously
• Memory Files: Creates persistent memory when given file access
• 65% Reduction: Less shortcut/loopholes behavior vs Sonnet 3.7

Industry Adoption

• GitHub: Integrating Sonnet 4 into GitHub Copilot
• Cursor: "State-of-the-art for coding"
• Rakuten: Validated 7-hour autonomous refactor
• Sourcegraph: "Substantial leap in software development"

New API Capabilities

4 new capabilities:

• Code execution tool
• MCP connector
• Files API
• Prompt caching (1 hour)

Claude Code Generally Available

• VS Code and JetBrains extensions (beta)
• GitHub Actions integration (demo)
• Claude Code SDK for custom agents
• GitHub PR integration via /install-github-app

Access

Already available via Anthropic API.

If you want to skip the new model restrictions, you can try it via Glama Gateway and OpenRouter.

So, is it hype?

Claude 4 models lead coding benchmarks and offer sustained performance for complex agent workflows. Opus 4 for maximum capability, Sonnet 4 for speed/cost balance. Both already available to test.

Source: Official Announcement

Will update this article to add interesting insights and facts as the day progresses.