Anthropic introduced the latest versions of its AI models, Claude Opus 4 and Claude Sonnet 4, which promise major improvements in coding, advanced reasoning, and building intelligent AI agents.

Meet the New Models

• Claude Opus 4 is now considered the best AI model in the world for coding. It performs strongly on complex, long tasks and agent workflows, making it ideal for developers and research teams.
• Claude Sonnet 4 is a big step up from the previous Sonnet 3.7, offering better coding, smarter reasoning, and more accurate responses to user instructions.

Smarter Thinking with Tools

Anthropic is also rolling out Extended Thinking with Tool Use (beta).

• Both new models can now use tools (like web search) during longer tasks.
• They can switch between thinking and using tools to give better results.
• They can even use several tools at once, follow instructions better, and if developers let them access local files, remember and reuse important information over time.

Claude Code Goes Public

After a successful preview, Claude Code is now available to everyone. It helps developers.

• Run background tasks with GitHub Actions
• Integrate directly into VS Code and JetBrains IDEs
• Suggest edits inside your code files, making it great for pair programming

New Developer Features via API

Anthropic’s API now includes.

• A code execution tool
• A multi-channel processing (MCP) connector
• A Files API
• The ability to cache prompts for up to 1 hour

Two Modes, One Goal: Better AI

Claude Opus 4 and Sonnet 4 offer.

• Fast responses for regular tasks
• Extended thinking mode for complex reasoning

They’re included in Anthropic’s Pro, Max, Team, and Enterprise plans, and Sonnet 4 is also available to free users.

You can use these models through the Anthropic API, Amazon Bedrock, and Google Cloud’s Vertex AI. Prices remain the same as before,

• Opus 4: $15 (input) / $75 (output) per million tokens
• Sonnet 4: $3 (input) / $15 (output) per million tokens

Why Opus 4 Is Special?

Claude Opus 4 stands out as the best model for software engineering.

• Top scores on the SWE bench (72.5%) and Terminal bench (43.2%)
• Handles long-running tasks for several hours
• Outperforms all previous models and Sonnet versions

Companies praise Opus 4.

• Cursor says it's a game-changer for coding and understanding large codebases.
• Replit reports major improvements across complex files.
• Block, Rakuten, and Cognition confirm its performance, especially for tough debugging and multi-hour automation tasks.

Sonnet 4: Powerful and Practical

Claude Sonnet 4 also performs impressively, especially in coding.

• Matches Opus 4 on SWE-bench (72.7%)
• Offers a balance of power and efficiency
• Ideal for both internal and customer-facing tools

Industry reaction.

• GitHub will use it in the new Copilot agent.
• Manus, iGent, Sourcegraph, and Augment Code note their precise reasoning, cleaner code, and fewer navigation mistakes in complex projects.

Better Safety and Smarter Memory

Both models now,

• Use fewer shortcuts and loopholes in tasks (65% improvement over Sonnet 3.7)
• Offer improved memory, especially Opus 4, which can write and keep memory files when allowed file access.

Example: Opus 4 played Pokémon Red, wrote its own navigation guide, and remembered key moves just like a real player would.

Also, Anthropic introduced thinking summaries, short versions of the model’s thought process. For advanced users, there's a new Developer Mode for full access.

Claude Code: Fully Integrated for Developers

Claude Code is now deeply integrated into development workflows.

• Available in VS Code, JetBrains, and in the terminal.
• Use the Claude Code SDK to create your own agents and apps.
• Try it on GitHub to respond to pull request feedback, fix errors, or change code.

Get Started Today

These models take a big step toward making AI your virtual teammate. They’re built with safety and responsible use in mind, even reaching AI Safety Level 3 (ASL-3).

Ready to try it? You can get started today on Claude, Claude Code, or via your favorite platform.