Anthropic has launched the new generation of AI models for its bot Claude. Claude Opus 4 and Claude Sonnet 4 present themselves as champions of code, "advanced reasoning," and agentic AI. This is a far cry from the usual flourishes of OpenAI, which readily communicates about its image generation tool with its assumed virality. ChatGPT also has more serious capabilities, but Claude is aimed more at professionals.
AI under close surveillance
And the new family of models aims to prove it. Claude Opus 4 is Anthropic's most powerful model; it is supposed to perform well for long-running tasks that require thousands of steps. It has the ability to run for several hours, allowing it to outperform other Sonnet models.
While not as powerful as its big brother Opus, Claude Sonnet 4 offers "considerable" improvements over the previous 3.7 model. It is more suited to everyday tasks.
Anthropic has also significantly reduced avoidance behaviors (use of shortcuts or loopholes) that are relatively common with Sonnet 3.7 when it is unsure how to complete a task. Both models are 65% less likely to exhibit this behavior on particularly vulnerable agent tasks.
When a developer grants Claude Opus 4 access to files stored on their computer, the model now maintains "memory files" to store key information. This ensures greater consistency and improved performance on agent tasks. The startup gives the example of a navigation guide for the Pokémon game: the agent records important information to improve its game!
These models therefore gain in savvy, and developers will certainly find relevant uses. But this is not without raising security questions as well. In 2023, Anthropic committed to not distributing models until it had developed security measures to restrict the most malicious uses.
Claude Opus 4 is launching with enhanced security measures dubbed "ASL-3" (AI Safety Level 3), the strictest ever implemented by Anthropic. Without this safeguard, AI is able to facilitate the access, production, or deployment of chemical, biological, or nuclear weapons for people with basic scientific knowledge! These measures include enhanced cybersecurity, protections against circumvention of restrictions, and additional systems to detect and reject certain dangerous behaviors. Phew.
These new models are available today for paid subscribers. Sonnet 4 is also available to free users.
Source: Anthropic
0 Comments