#opus4

Claude 4 models

The latest AI models from Amazon-backed AI safety and research business Anthropic, founded by former OpenAI researchers, are Claude Opus 4 and Claude Sonnet 4. These models launched on May 22 and 23, 2025, setting “new standards for coding, advanced reasoning, and AI agents“. Anthropic stopped investing in chatbots last year to focus on improving Claude's ability to handle complex tasks like research and coding, so the launch is a bold move away from chatbots to become a well-known AI coding platform.

Opus 4Claude

They call Claude Opus 4 the “best coding model in the world” and Anthropic's most powerful. In customer testing, it ran virtually a full workday (seven hours) independently and performed well on complex, time-consuming activities and agent processes. For complex use scenarios requiring “frontier intelligence,” use Opus 4:

Intelligent AI.

Full-stack app development and codebase refactoring.

Research synthesis, agentic search, and deep research.

Long-term independent work that prioritises expertise and precision.

Content development focusses on natural writing and humanity.

Fourth Claude Sonnet

Anthropic's mid-size Claude Sonnet 4 balances pricing and performance. Though it replaced Claude Sonnet 3.7, it is much better. It has “superior coding and reasoning,” more accurate responses, and 65% less “reward hacking.” Claude Sonnet 4 suits high-volume, broad activities like:

Fixing bugs and reviewing code.

AI assistants for real-time client communications.

Condensing market signals or dashboards requires good research.

Massive material generation and analysis.

Being a task-specific subagent in multi-agent systems.

Claude Sonnet 4 and Opus 4 are hybrid reasoning models that can answer questions quickly and allow for deeper reasoning through “extended thinking” mode. With more time to think about solutions, this prolonged thinking mode helps models perform better on difficult tasks. The models can summarise their logic in a “user-friendly” way. A new Developer Mode lets you see full thought sequences before processing.

Both models can now leverage parallel tools like online search to contact several APIs or plugins at once to speed up procedures and reduce errors. They can extract and store relevant information in local files to construct “memory files” or “tacit knowledge” over time, boosting continuity and reliability on long-term activities. The models also follow directions better.

Opus 4 and Claude Sonnet 4 scored industry-leading on the SWE-bench coding benchmark, highlighting the focus on coding. Opus 4 also performs well on Terminal Bench real-world coding tests. Performance fluctuates, therefore internal benchmarks should be taken “with a grain of salt.” In high school maths, benchmarks show regressions compared to earlier models. Even while AI models struggle to write high-quality software and sometimes make mistakes, their potential to boost productivity is driving their rapid adoption.

To help developers, Anthropic has made its Claude Code agentic command-line tool public. With Claude Code's interaction with GitHub, VS Code, and JetBrains, file edits are visible quickly. The extensible Claude Code SDK lets you create unique agents and apps. A code execution tool, MCP connection, Files API, and prompt caching are new API features.

Opus 4 and Claude Sonnet 4 are accessible via Google Cloud Vertex AI, Amazon Bedrock, and the Anthropic API. Databricks customers can use them natively. Opus 4 is part of Anthropic's premium Claude plans (Pro, Max, Team, and Enterprise), while Claude Sonnet 4 is free and paid. Claude Sonnet 4 starts at $3 per million input tokens and $15 per million output tokens, while Opus 4 starts at $15 and $75. Timely caching and batch processing save money.

Anthropic tested and assessed the models with additional experts to ensure safety, security, and dependability. They come with improved cybersecurity and harmful content identification. Opus 4 may “substantially increase” the capacity of a STEM-trained person to acquire, produce, or deploy chemical, biological, or nuclear weapons, according to internal testing. The models are evaluated against Anthropic's “ASL-3” model

Early clients like Palo Alto Networks, Replit, Cursor, Rakuten, Augment Code, and others have observed improvements in agent performance, difficult job management, coding pace, and code quality. Claude on Vertex AI accelerated Palo Alto Networks code development by 20%–30%. Opus 4 was the first to increase code quality while debugging and editing without slowing down, Block said.

Anthropic reported $2 billion in annualised sales in the first quarter of 2025, more than double the previous quarter. The company targets $12 billion in 2027. Wall Street is still investing, and Anthropic has received billions from Amazon and has a $2.5 billion credit line to cover rising development costs.