On May 22, 2025, Anthropic announced the next generation of its Claude models: Claude Opus 4 and Claude Sonnet 4. The company positions the two models as steps forward in coding, long-running reasoning, and AI agent workflows.
Capabilities and differences
-
Claude Opus 4 is presented as Anthropic’s most powerful and the world’s best coding model. It shows sustained performance on long tasks requiring focused work over thousands of steps and can operate continuously for several hours. Reported benchmark results include 72.5% on SWE-bench and 43.2% on Terminal-bench. Anthropic cites multiple partner tests—such as a 7-hour autonomous open-source refactor run by Rakuten—that demonstrate its sustained performance and ability to improve code quality during editing and debugging.
-
Claude Sonnet 4 is a significant upgrade from Sonnet 3.7, delivering industry-leading coding performance (SWE-bench 72.7%) with improved steerability for more precise instruction following. While it generally does not outperform Opus 4 across the board, Anthropic frames Sonnet 4 as an efficient, high-performance option that is also available to free users.
Multiple external organizations provided early feedback or tests—GitHub, Replit, Cursor, Block, Cognition, Manus, iGent, Sourcegraph, Augment Code and others—highlighting improvements in multi-file changes, agentic scenarios, instruction following, and reduced navigation errors in codebases.
Extended thinking and tool use (beta)
Both models can now use tools (for example, web search) during an "extended thinking" mode. In this mode the model alternates between internal reasoning and invoking tools to improve answers. The models can execute tools in parallel. Anthropic reports a 65% reduction in the tendency to use shortcuts or loopholes on agentic tasks compared with Sonnet 3.7.
Memory and local file access
Opus 4 shows substantial gains in memory capabilities when given local file access by developers: it can create and maintain "memory files" to extract and store key facts, improving long-term task awareness, coherence, and agent performance. Anthropic illustrates this with examples where Opus 4 recorded notes (e.g., while playing Pokémon) that helped sustain performance over time.
Claude Code becomes generally available
Claude Code exits research preview and is now generally available. Key features:
- Background task support via GitHub Actions.
- Native VS Code and JetBrains extensions: Claude’s suggested edits appear inline in files, streamlining review and pair programming workflows.
- An extensible Claude Code SDK to build custom agents and applications around the same core agent.
- A beta "Claude Code on GitHub" integration that can be tagged on PRs to respond to reviewer feedback, fix CI errors, or modify code.
New API capabilities
Anthropic is releasing four new API capabilities aimed at enabling more powerful AI agents for developers:
- A code execution tool
- An MCP connector
- A Files API
- Prompt caching for up to one hour
Availability and pricing
Both Claude Opus 4 and Sonnet 4 are hybrid models with two modes: near-instant responses and extended thinking for deeper reasoning. The Pro, Max, Team and Enterprise Claude plans include both models and extended thinking; Sonnet 4 is also available to free users. The models are accessible via Anthropic’s API, Amazon Bedrock and Google Cloud Vertex AI. Pricing remains the same as prior Opus and Sonnet releases: Opus 4 at $15/$75 per million tokens (input/output) and Sonnet 4 at $3/$15.
Benchmarks and methodology
Anthropic reports that the Claude 4 family leads on SWE-bench Verified, a benchmark for real software engineering tasks. The announcement provides results across several benchmarks—GPQA Diamond, MMMLU, MMMU, AIME and TAU-bench—and notes whether extended thinking was used. Examples of reported scores:
- SWE-bench: Opus 4 72.5%, Sonnet 4 72.7%
- Terminal-bench: Opus 4 43.2%
- Using a high-compute internal selection process (parallel attempts and scoring), Anthropic reports scores of 79.4% and 80.2% for Opus 4 and Sonnet 4 respectively.
The post also documents methodological choices for TAU-bench and SWE-bench, such as increasing the maximum number of model completions from 30 to 100 for extended trajectories and removing the third planning tool used in Claude 3.7 Sonnet for the Claude 4 evaluations.
Safety, testing and next steps
Anthropic emphasizes extensive testing and evaluation to reduce risk and increase safety, including implementing measures for higher AI Safety Levels like ASL-3. For users who require raw chains of thought for advanced prompt engineering, Anthropic offers a Developer Mode (contact sales) to retain full access to internal reasoning traces.
Getting started
Anthropic positions these models as a step toward a virtual collaborator that maintains context and focus on longer projects. Claude, Claude Code and the models on partner platforms are available now; Anthropic encourages users to try them and provide feedback to guide further improvements.
