Anthropic Introduces Claude 3.7 Sonnet

Anthropic just released their latest, greatest model—Claude 3.7 Sonnet.

Claude 3.7 Sonnet, says Anthropic, is the first "hybrid reasoning model" on the market. Which is a fancy way of saying this model operates in two modes: a quick-response mode and an “extended thinking” mode that can showcase its step-by-step logic.

Anthropic believes the ability to quickly fire off answers or deeply think through complex tasks shouldn’t require two separate models—instead, a single system should handle both, just like a human brain.

Sound ambitious? It is. And early testers claim Claude 3.7 Sonnet backs up that ambition, especially when it comes to coding. Multiple companies report that this release outperforms previous models, particularly in building complex apps and working on full-stack projects.

So what’s going on here, and how does this fit into the bigger race toward more capable AI?

I got the scoop from Marketing AI Institute founder and CEO Paul Roetzer on Episode 138 of The Artificial Intelligence Show.

What Makes It Different?

In standard mode, Claude 3.7 Sonnet behaves like a typical large language model, delivering responses at high speed. But then it also has an “extended thinking” mode which gives the AI time to reflect step-by-step. 

You can toggle between the two. You can get a quick, lightweight answer from Claude. Or instruct it to slow down and think methodically before responding. If you choose the latter, the model will take more time to work through a problem, showing you its reasoning as it goes.

API users can also decide exactly how much “thinking” the model does—right down to how many tokens it’s allowed to spend on reasoning. If you’re strapped for time (or cash), you can limit the budget. If accuracy is paramount, give it the freedom to think for a while.

The company hopes that this one integrated AI “brain” will feel more natural to users and yield better results across tasks—from casual Q&A to hardcore programming.

But, buyer beware, they're also hyping it up a lot, says Roetzer.

Make no mistake, the model is impressive. But all the major labs are doing this as well. Everyone is working on some kind of “reasoning plus LLM” hybrid under the hood. So while Anthropic might beat others to some limited release, we should expect more models to have these same capabilities.

Where This Is Headed

But the real story isn't where Claude is today. (Though you should go try out Claude 3.7 Sonnet.) It's where Claude is going.

As part of the Claude 3.7 Sonnet release post, Anthropic laid out a timeline for the model's evolution. In 2024, Claude was all about assistance, or helping you do your current work better. In 2025, says Anthropic, it'll be much more about collaboration, or doing "hours of independent work for you."

But 2027 is really when things get interesting. Then, according to the company, is when Claude becomes a pioneer, finding "breakthrough solutions to challenging problems that would have taken teams years to achieve."

This is basically an assumption we'll hit artificial general intelligence (AGI) in 2027, says Roetzer.

"The assumption from all these major labs seems to be, we get AGI. We enter the phase where these things are just now better than humans at basically all cognitive tasks."

The timeline, he notes, increasingly converges around 2027 as the major labs talk more about AGI and make predictions. Which means it's really, really important to pay attention.

Because some of the top minds in AI think we've got just a couple years before it happens.

What's Next for Anthropic?

As we hurtle towards AGI, there's also an open question of what happens to Anthropic.

They're rumored to be raising another $3.5 billion, which would value the company at $61.5 billion. That's not a small amount of money, but it still puts them well behind heavyweights like OpenAI and Google.

"I think there's a chance Anthropic manages to stay independent and achieve their research and business missions and get to AGI," says Roetzer.

But he thinks there's a greater probability the company takes a different path.

The most likely is: They get acquired. They're building great models. They have an amazing research team. And they seemingly have a greater focus on safety and alignment than many other players. 

But, says Roetzer, they lack other advantages that may be needed win the AI arms race.

"They have no data," he says, meaning any proprietary data to train on like other platforms (Google, Meta, xAI) have. "They have no products you would get data from."

They also lack the distribution advantages of some rivals.

"They have no distribution other than the app itself," he says. "So they're a little bit behind there."