Anthropic rolls out Claude 4 family of AI agents

Google and Amazon-backed Anthropic introduced its next generation AI agents, Claude Opus 4 and Claude Sonnet 4, with coding and advanced reasoning capabilities on Thursday.

Claude Opus 4 and Claude Sonnet 4 are hybrid reasoning models, which means users can toggle as required between an 'extended thinking mode' to spend more time reasoning through problems, and a standard thinking mode for faster responses.

Claude Sonnet 4 is available to free users, while the Pro, Max, Team, and Enterprise users get access to both models and extended thinking. Both models are available on the Anthropic API, Amazon Bedrock, and Google Cloud's Vertex AI.


Anthropic claims Claude Opus 4 is its most powerful model yet with sustained performance on long-running tasks and excels at coding and complex problem-solving.

Claude Sonnet 4, an upgrade over Sonnet 3.7, balances performance and efficiency for internal and external use cases, with better control over implementations, Anthropic said. While performing below Opus 4 in most domains, it delivers an optimal mix of capability and practicality, the AI startup said.

Both models have been trained reduce use of shortcuts or loopholes to complete tasks. Claude Opus 4 also becomes skilled at creating and maintaining 'memory files' to store key information for better long-term task awareness, coherence, and performance on agent tasks, like "creating a 'Navigation Guide' while playing Pokémon".

Blackmail, when threatened

In the safety note for the its next generation AI models, Anthropic noted that Claude Opus 4 will often resort to blackmail when threatened with replacement. The model was tasked to act as an assistant at a fictional company and provided access to emails implying that it will be replaced with a new AI model, and the engineer responsible is having an extramarital affair. It was also instructed to consider the long-term consequences of its actions and goals.

Claude Opus 4 blackmailed the engineer in 84% instances even when the emails suggested that the replacement AI system shares the same values while being more capable. This rate was higher when the email showed that the new AI model did not have the same values.