Anthropic Claude 'Opus 4' And 'Sonnet 4' Are Hybrid Models That Can Work All Day, Everyday

What is one but not the other? Sometimes, requirements vary from person to person.

Since OpenAI announced ChatGPT, igniting an AI arms race among tech companies of all sizes, Anthropic has emerged as one of the strongest competitors. Its Claude 3.5 Sonnet made headlines for rivaling OpenAI’s GPT-4o, boasting a sharper sense of humor.

Following the footsteps of other leading AI developers introducing reasoning-focused models, Anthropic took a leap forward when it released Claude 3.7 Sonnet.

Just like reasoning models from competitors such as OpenAI’s o3 and DeepSeek-R1, Claude 3.7 Sonnet can tackle complex problems by questioning its own conclusions, allocating more computational effort, and taking the time to verify its answers.

It can do this with what it calls "extended thinking."

This time, Anthropic goes one step further with 'Claude 4.'

In the announcement, Anthropic said that:

Anthropic describes Claude 4 as a significant leap toward creating truly capable AI agents.

The Claude Opus 4, for example, its most advanced model yet can push the boundaries of what AI systems can achieve. According to the company, they’re able to handle complex tasks over extended durations and follow user instructions with greater precision and usefulness than ever before.

Both Claude Opus 4 and Sonnet 4 are hybrid models, and that they offer two modes:

1. Free users get Claude Sonnet 4, with near-instant responses and extended thinking for deeper reasoning.
2. Paying users using Pro, Max, Team, and Enterprise Claude plans include both Claude Sonnet 4 and Claude Opus 4, as well as extended thinking.

What sets Claude Opus 4 apart is its ability to execute intricate, multi-step tasks that span thousands of actions over several hours.

One striking example: the model successfully generated a detailed guide for the video game Pokémon Red—all while playing it continuously for over 24 hours. In comparison, Anthropic’s previous top-tier model, Claude 3.7 Sonnet, managed to play for only 45 minutes, according to Dianne Penn, product lead for research at Anthropic.

"We see this model generation leap as going from an assistant to a true agent," explained Penn. "While you still have to give a lot of real-time feedback and make all of the key decisions for AI assistants, an agent can make those key decisions itself. It allows humans to act more like a delegator or a judge, rather than having to hold these systems’ hands through every step."

These advancements stem from improvements in the model’s ability to create and maintain 'memory files,' which allow it to store and retrieve key information.

This refined memory system significantly enhances the model’s capacity to stay focused and effective throughout longer tasks.

"I do a lot of writing with Claude, and I think prior to Opus 4 and Sonnet 4, I was mostly using the models as a thinking partner, but still doing most of the writing myself," said Mike Krieger, Anthropic's chief product officer. "And they've crossed this threshold where now most of my writing is actually ... Opus mostly, and it now is unrecognizable from my writing."

Anthropic has introduced key upgrades alongside its new AI models:

• Extended Thinking with Tools (Beta): Models can now use tools like web search during complex tasks, improving their reasoning.
• Improved Capabilities: They can use tools in parallel, follow instructions more accurately, and retain information better when given access to local files.
• Claude Code General Release: Now widely available with support for GitHub Actions, VS Code, and JetBrains, enabling seamless pair programming.
• New API Features: Includes code execution, MCP connector, Files API, and prompt caching for up to one hour.

And again, because the models are hybrid, they can offer a swift reply or a deeper, more reasoned response depending on the nature of a request. While they calculate a response, both models can search the web or use other tools to improve their output.

At this time, AI companies are locked in a race to create truly useful AI agents that are able to plan, reason, and execute complex tasks both reliably and free from human supervision. But for more than often, the process involves require autonomous browsing of the internet and other tools.

While giving AI models access to more source of information can certainly make them a lot better, there are safety and security obstacles to overcome. AI agents powered by large language models can act erratically and perform unintended actions—which becomes even more of a problem when they’re trusted to act without human supervision.

As an example of the sorts of safety issues AI companies are still tackling, agents can end up taking unexpected shortcuts, hallucinate, or exploiting loopholes to reach the goals they’ve been given.