Introducing Claude 4: Summary

Overview

Anthropic has announced the next generation of Claude models: Claude Opus 4 and Claude Sonnet 4, released on May 22, 2025. These models set new standards for coding, advanced reasoning, and AI agents.

Key Models & Performance

• Claude Opus 4: The world's best coding model, designed for sustained performance on complex, long-running tasks and agent workflows.

  • Benchmark Leadership: Leads on SWE-bench (72.5%) and Terminal-bench (43.2%).
  • Endurance: Can work continuously for several hours, outperforming all previous Sonnet models.
  • Partner Feedback:
    • Cursor: "State-of-the-art for coding and a leap forward in complex codebase understanding."
    • Rakuten: Validated with a 7-hour open-source refactor with sustained performance.
    • Cognition: "Excels at solving complex challenges that other models can't."

• Claude Sonnet 4: A significant upgrade to Sonnet 3.7, balancing performance and efficiency.

  • Benchmark Leadership: Achieves state-of-the-art 72.7% on SWE-bench.
  • Partner Feedback:
    • GitHub: Will power the new coding agent in GitHub Copilot.
    • iGent: Reduced navigation errors from 20% to near zero.
    • Sourcegraph: "A substantial leap in software development."

New Capabilities & Features

• Extended Thinking with Tool Use (Beta): Both models can use tools (e.g., web search) during extended thinking, alternating between reasoning and tool use.
• New Model Capabilities:
  • Parallel tool execution.
  • More precise instruction following.
  • Improved Memory: When given local file access, models can extract and save key facts to maintain continuity over time. Opus 4 excels at this, creating "memory files" for long-term task awareness.
• Reduced Shortcut Behavior: Both models are 65% less likely to use shortcuts or loopholes than Sonnet 3.7 on susceptible agentic tasks.
• Thinking Summaries: A smaller model condenses lengthy thought processes (~5% of the time). Full chain-of-thought access is available via a new "Developer Mode."

Claude Code & Developer Tools

• Claude Code is Generally Available:
  • New IDE Extensions (Beta): Native integrations for VS Code and JetBrains, displaying edits inline.
  • Background Tasks: Supports GitHub Actions.
  • Claude Code SDK: Allows building custom agents and applications.
  • Claude Code on GitHub (Beta): Tag Claude Code on PRs to respond to feedback, fix CI errors, or modify code.
• New API Capabilities:
  • Code execution tool
  • MCP connector
  • Files API
  • Prompt caching for up to one hour

Availability & Pricing

• Availability: Both models are available on the Claude platform, Amazon Bedrock, and Google Cloud's Vertex AI.
• Plans: Pro, Max, Team, and Enterprise plans include both models and extended thinking. Sonnet 4 is also available to free users.
• Pricing (per million tokens):
  • Opus 4: $15 (input) / $75 (output)
  • Sonnet 4: $3 (input) / $15 (output)

Appendix: Technical Details

• Benchmark Methodology:
  • SWE-bench: Uses a simple scaffold with bash and file editing tools. "High compute" scores (Opus 4: 79.4%, Sonnet 4: 80.2%) use parallel attempts and rejection sampling.
  • TAU-bench: Used extended thinking with tool use and increased the maximum step limit from 30 to 100.
• Extended Thinking Scores: Benchmarks like GPQA Diamond, MMMLU, and AIME report highest scores with extended thinking. Scores without extended thinking are also provided.