Claude 4 release notes / what's new

Anthropic has introduced the next generation of Claude models, Claude Opus 4 and Claude Sonnet 4. These models set new benchmarks for coding, advanced reasoning, and AI agents, introducing hybrid modes that allow users to switch between near-instant responses and extended thinking for deeper reasoning.

Model Performance and Benchmarks

Claude Opus 4 and Claude Sonnet 4 lead the industry in software engineering and reasoning tasks. Opus 4 is positioned as the world's best coding model, while Sonnet 4 provides a balance of performance and efficiency.

Coding and Software Engineering

  • SWE-bench: Claude Opus 4 achieved 72.5% and Claude Sonnet 4 achieved 72.7% pass@1.
  • Terminal-bench: Claude Opus 4 leads with 43.2%, significantly outperforming other models.
  • Agentic Terminal Coding: When using Claude Code as the agent framework, Opus 4 reaches 43.2% and Sonnet 4 reaches 35.5%.

Reasoning and Knowledge

  • GPQA Diamond: Opus 4 scored 79.6% (with 83.3% using parallel test-time compute).
  • AIME 2025: Opus 4 achieved 75.5% (increasing to 90.0% with parallel test-time compute).
  • MMMLU: Opus 4 scored 88.8% and Sonnet 4 scored 86.5%.
  • MMMU (Validation): Opus 4 scored 76.5% and Sonnet 4 scored 74.4%.

Key Technical Capabilities

Claude 4 introduces several architectural and functional improvements designed to enhance the autonomy and reliability of AI agents.

Extended Thinking and Tool Use

Both models now support extended thinking with tool use (beta). This allows the models to alternate between internal reasoning and the execution of tools—such as web search—to refine their responses. Additionally, the models can now execute tools in parallel.

Memory and Continuity

Claude Opus 4 demonstrates significantly improved memory capabilities. When granted access to local files by developers, the model can create and maintain "memory files" to extract and save key facts. This allows the model to build tacit knowledge and maintain continuity over long-term tasks.

Reliability and Precision

Anthropic has reduced the tendency of models to use shortcuts or loopholes to complete tasks. Both Opus 4 and Sonnet 4 are 65% less likely to engage in this behavior on agentic tasks compared to Sonnet 3.7.

Thinking Summaries

To manage lengthy thought processes, Claude 4 models use a smaller model to condense thinking summaries. This is applied in approximately 5% of cases where thought processes are too long to display in full.

Claude Code and Developer Ecosystem

Claude Code is now generally available, expanding the developer's ability to collaborate with the AI in the terminal and IDE.

  • IDE Integration: New beta extensions for VS Code and JetBrains allow proposed edits to appear inline within files.
  • Claude Code SDK: An extensible SDK allows developers to build custom agents. A beta integration for GitHub allows Claude Code to respond to PR feedback, fix CI errors, and modify code via tagging.
  • API Enhancements: Four new API capabilities have been released: a code execution tool, MCP connector, Files API, and prompt caching for up to one hour.

Industry Adoption and Feedback

Several frontier AI companies and platforms have integrated Claude 4:

  • GitHub: Will introduce Claude Sonnet 4 as the model powering the new coding agent in GitHub Copilot.
  • Cursor: Describes Opus 4 as a leap forward in complex codebase understanding.
  • Replit: Reports improved precision for complex changes across multiple files.
  • Block: Notes that Opus 4 is the first model to boost code quality during editing and debugging in its agent, codename goose.
  • Rakuten: Validated Opus 4's ability to run an open-source refactor independently for 7 hours.
  • Cognition: States that Opus 4 excels at solving complex challenges that previous models missed.

Availability and Pricing

Claude Opus 4 and Sonnet 4 are available via the Claude.ai interface, Amazon Bedrock, and Google Cloud Vertex AI.

  • Pricing:
    • Opus 4: $15 per million input tokens / $75 per million output tokens.
    • Sonnet 4: $3 per million input tokens / $15 per million output tokens.
  • Access: Pro, Max, Team, and Enterprise plans include both models. Sonnet 4 is also available to free users.

Sources

Related