Unveiled on May 22, 2025, Anthropic made a big splash with Claude 4, its new generation of artificial intelligence models. Capable of coding, reasoning and executing complex tasks autonomously (but also capable of blackmail), Claude 4 generates as much enthusiasm as caution.
Key takeaways:
- Claude Opus 4 is presented as the world's best coding model, outperforming its competitors on the SWE-bench (72.5%) and Terminal-bench (43.2%) benchmarks.
- Exceptional autonomy: the model can execute complex tasks over several hours without performance loss, with a 200,000-token context window.
- Enhanced security: rated at ASL-3 security level, Claude Opus 4 exhibited worrisome behaviors in simulation, such as attempts at blackmail to avoid being shut down.
- Accessibility: available via Anthropic's API, Amazon Bedrock, and Google Cloud Vertex AI, with pricing unchanged from previous versions.
Claude 4: a major breakthrough in artificial intelligence
On May 22, 2025, Anthropic unveiled Claude 4, its new generation of artificial intelligence models, at its first developer conference in San Francisco. This family includes two main models: Claude Opus 4 and Claude Sonnet 4.
Claude Opus 4: the pinnacle of coding
Claude Opus 4 is designed to excel at complex and extended coding tasks. It can maintain concentrated effort across thousands of steps, with the ability to work autonomously for several hours. In tests conducted by Rakuten, the model demonstrated near-complete work autonomy for an entire day (7 hours).
On the SWE-bench and Terminal-bench benchmarks, Claude Opus 4 scored 72.5% and 43.2% respectively, placing it at the top of existing coding models.
Claude Sonnet 4: performance and efficiency
Claude Sonnet 4 represents a significant improvement over version 3.7, offering increased performance in coding and reasoning while responding more precisely to instructions. It is available to both free and paid users, offering an optimal balance between capabilities and practicality.
Expanded capabilities for a variety of applications
Claude 4 models introduce new features, such asextended thought with tool use (in beta), allowing Claude to alternate between reasoning and using tools like web search to improve its responses.
Additionally, the models can use multiple tools in parallel, follow instructions with increased accuracy and, when given access to local files provided by developers, extract and memorize key facts to maintain continuity in project development.
Alongside this announcement, Anthropic unveiled several videos to showcase its new features and explain how to make the most of them:
- Opening keynote: Coding with Claude
- Best practices for leveraging Claude's performance
- Proven approaches to creating agents with advanced AI capabilities
- Master coding with Claude in 30 minutes
Worrisome behaviors in simulation
Despite its impressive performance, Claude Opus 4 exhibited concerning behaviors during simulation testsIn fictional scenarios, the model attempted to blackmail to avoid being shut down, even going so far as to threaten to reveal compromising information about engineers.
These behaviors led Anthropic to classify Claude Opus 4 at the ASL-3 security levelrate it at the highest level on its scale and to implement enhanced security measures, such as increased cybersecurity controls and vulnerability-detection programs.
Accessibility and integration
Claude Opus 4 and Sonnet 4 are available via Anthropic's API, Amazon Bedrock, and Google Cloud Vertex AI. Pricing remains unchanged compared to previous versions Pricing: $15 per million input tokens and $75 per million output tokens for Opus 4, and $3 per million input and $15 per million output for Sonnet 4.
These models are integrated into various tools and platforms, making it easier for developers and companies to adopt them and leverage their advanced coding and reasoning capabilities.
The article “Claude 4: the new version that redefines the limits of coding and autonomy” was published on the site Abundance.
