Anthropic Unveils AI Model with Customizable Thinking Duration

# Claude

Anthropic's latest AI model, Claude 3.7 Sonnet, is shaking things up in the world of AI. Dubbed the industry's first "hybrid AI reasoning model," it's designed to give you both quick answers and more thoughtful responses, depending on what you need. You can even choose how long you want Claude to "think" about your questions—pretty cool, right?

The idea behind Claude 3.7 Sonnet is to make using AI a lot simpler. Instead of having to pick from a bunch of different models, Anthropic wants one model to handle everything. Starting Monday, everyone can get their hands on Claude 3.7 Sonnet, but if you want to use its fancy reasoning features, you'll need to sign up for one of Anthropic's premium plans. Free users will still get a solid version of the model, which Anthropic says is better than their last one, Claude 3.5 Sonnet.

Now, about the cost—Claude 3.7 Sonnet will set you back $3 per million input tokens and $15 per million output tokens. That's a bit pricier than some other models out there, like OpenAI's o3-mini and DeepSeek's R1, but remember, those are just reasoning models, not hybrids like Claude.

Claude 3.7 Sonnet is Anthropic's first model that can "reason," which is a big deal in the AI world. Reasoning models take a bit more time and computing power to answer questions, breaking them down into smaller steps for more accurate answers. It's not exactly like human reasoning, but it's modeled after it.

Anthropic's got big plans for Claude. Eventually, they want it to figure out how long it should "think" about questions on its own, without you having to set it up. Dianne Penn, Anthropic's product and research lead, told TechCrunch they see reasoning as just one of the many things a frontier model should be able to do, all rolled into one.

With Claude 3.7 Sonnet, you'll even get to see how it plans out its answers through a "visible scratch pad." Most of the time, you'll see the whole process, but Anthropic might redact some parts for safety reasons.

Anthropic says they've fine-tuned Claude's thinking modes for real-world tasks, like tough coding problems or tasks that need a bit of autonomy. If you're a developer using Anthropic's API, you can even control how much "thinking" Claude does, balancing speed and cost with the quality of the answer.

In some tests, Claude 3.7 Sonnet did pretty well. On SWE-Bench, a test for real-world coding tasks, it scored 62.3%, beating out OpenAI's o3-mini, which got 49.3%. And on TAU-Bench, which measures how well an AI can interact with simulated users and external APIs in a retail setting, Claude 3.7 Sonnet scored 81.2%, compared to OpenAI's o1 model, which got 73.5%.

Anthropic also says Claude 3.7 Sonnet will be less likely to refuse to answer questions, making more nuanced distinctions between harmful and benign prompts. They've reduced unnecessary refusals by 45% compared to Claude 3.5 Sonnet.

Along with Claude 3.7 Sonnet, Anthropic's also rolling out a new tool called Claude Code. It's a research preview that lets developers run specific tasks through Claude right from their terminal. In a demo, Anthropic showed how Claude Code can analyze a coding project with a simple command like, "Explain this project structure." You can even use plain English to modify a codebase, and Claude Code will describe its edits as it goes, test for errors, and push the project to a GitHub repository.

Claude Code will be available to a limited number of users on a "first come, first serve" basis.

Anthropic's releasing Claude 3.7 Sonnet at a time when AI labs are churning out new models left and right. While Anthropic's usually taken a more careful, safety-focused approach, this time they're looking to lead the pack. But how long they'll stay ahead is anyone's guess—especially with OpenAI's CEO, Sam Altman, saying they've got a hybrid AI model of their own coming in "months."

Related article

Comments (50)