Anthropic intros next generation of Claude AI models

Claude 4 models

Claude Opus 4 is a coding-specific model that can sustain performance on complex, long-running tasks and agent workflows. In testing Opus 4, the model demonstrated about seven hours of fully autonomous coding work, according to Anthropic.

Claude Sonnet 4 replaces Claude 3.7 Sonnet and includes coding and reasoning capabilities.

"What we're trying to do with these models is just to continue to push along the frontier," said Alex Albert, head of developer relations at Anthropic. "With Claude 4, and with Opus 4 in particular, we want to be setting new standards for coding, for advanced reasoning generally and for agents."

Other than the models, Anthropic introduced new capabilities. For example, new extended thinking with tool use means both models can use tools like web search while performing deep thought. Instead of sharing Claude's thinking process, the models share thinking summaries.

The vendor also released four new capabilities on the Anthropic API that enable developers to build more powerful AI agents. The features are a code execution tool, a connector to Anthropic's widely adopted Model Context Protocol, Files API and the ability to cache prompts for up to an hour, Anthropic said.

Claude Opus 4 and Sonnet 4 are hybrid models that can provide either thinking or instant responses. The models are available to all paid subscribers, and free subscribers can access Claude Sonnet 4.

Both models are available on the Anthropic API, Amazon Bedrock and Google's Vertex AI. Anthropic's new generation of models arrives more than a year after the foundation model provider introduced the previous generation of Claude and three months after introducing Claude 3.7 Sonnet and Claude Code.

Anthropic's strategy

Anthropic's strategy for its updated models shows that the independent GenAI vendor is focusing not just on meeting benchmarks, but also on solving real-world problems, "with transparency and observability and understandability in terms of how the models are making decisions," said Bradley Shimmin, an analyst at The Futurum Group.

We are going to start to see these models doing far more than simple code completion and code documentation.

He added that the developer community has also recognized for the past few months to a year that the Claude models are adept at engineering code. Therefore, Opus 4's ability to read code for a lengthy period marks a new beginning in the GenAI market.

"We are going to start to see these models doing far more than simple code completion and code documentation," Shimmin said. "They're going to be used in addition to that to help companies solve some very perplexing problems they historically have always run into with regard to maintaining complex code."

This means that more applications could emerge from models like Opus 4 than just basic chain-of-thought reasoning, he said.

Another significant marker of Opus 4 is the ability to switch between "thinking mode" and "tool usage mode," said Andy Thurai, CEO and founder of The Field CTO.

"This is a significant switch from a single dialogue context model to intelligent agents that can switch based on need," Thurai said.

He added that thinking summaries could be useful "when AI is used for subjective decision-making, and the thinking behind the decision-making needs to be captured."

A bigger trend

These moves also show a trend with Anthropic and other model makers, such as Google with its Gemini model line, in which the models are not only performing tasks, but also becoming collaborators.

"We're starting to see less of a 'help me code something' and more of a 'help me to design or to understand the problem' -- to conceptualize a solution, to design that solution and to at least get started on the implementation of it," Shimmin said.

Compared with competitors like OpenAI and Google, which issue more frequent releases, Anthropic seems to take what some might consider a measured approach with its model releases, he added.

"That might be OK for a company like Anthropic because what they're trading on isn't keeping mindshare," Shimmin said. "Instead, it's the word of mouth and preference amongst those who influence buying decisions."

Esther Shittu is an Informa TechTarget news writer and podcast host covering AI software and systems.