Artificial intelligence model maker Anthropic PBC has thrown down the gauntlet to OpenAI, DeepSeek Ltd. and others in the industry with today’s launch of a new frontier model called Claude 3.7 Sonnet.

Unlike its previous models, Claude 3.7 Sonnet can “think” about questions for as long as the users ask it too, so depending on how long it considers things, its responses could be vastly different.

The startup says Claude 3.7 Sonnet is the first “hybrid AI reasoning model” because it’s capable of answering in real time, or it can generate better-thought-out responses if desired. Users can choose when to activate its reasoning capabilities, and then specify how long they want it to consider the question.

Claude 3.7 Sonnet is being made available today to everyone, including free users, but only those who pay for a premium subscription will get access to its advanced reasoning features. Free users will only get the real-time version, though the company says it’s still an improvement on its predecessor, Claude 3.5 Sonnet.

The company said Claude 3.7 Sonnet will cost $3 per 1 million input tokens, which means you could enter about 750,000 words (more than the entire “Lord of the Rings” trilogy) for just $3. It also charges $15 per 1 million output tokens.

As such, Claude 3.7 Sonnet is more expensive than OpenAI’s o3-mini reasoning model and DeepSeek’s R1, which cost about three and six times less. That said, Anthropic’s models have always been more costly, with users paying exactly the same rates to access Claude 3.5 Sonnet. So they’re getting the new reasoning features without paying anything extra.

Claude 3.7 Sonnet represents the company’s first reasoning model, which uses more computing power and takes more time to generate responses than traditional models. They work by breaking down the user’s question or problem into a series of small steps, considering each of them individually before compiling their response, and the technique often results in a better answer.

For now, users have to choose how long Claude 3.7 Sonnet will think about a question themselves. But in a forthcoming update, the company says, the model will be able to determine the most suitable timeframe for thinking by itself, striking an optimal balance between cost and answer quality.

Anthropic’s product and research chief Dianne Penn told VentureBeat in an interview that the aim is to get the model to know when an instant answer is needed, and when a more considered response is appropriate. “The model itself should recognize when a problem requires more intensive thinking and adjust, rather than requiring users to explicitly select different reasoning modes,” she said.

Another cool feature of Claude 3.7 Sonnet is that it will show its internal thinking processes through a “visible scratch pad.” Users will be able to see the entire chain of thought for most prompts, though in some cases it may redact certain elements for trust and safety considerations, Penn said.

As for its performance, Claude 3.7 Sonnet stands up well against its competitors, scoring 62.3% on the real-world coding benchmark SWE-Bench, compared with 49.3% for OpenAI’s o3-mini and 49.2% for DeepSeek R1.

On another test designed to measure its ability to interact with simulated users and external application programming interfaces, called TAU-Bench, Claude 3.7 Sonnet scored 81.2%, surpassing OpenAI’s o1 model’s score of 73.5%.

The startup adds that Claude 3.7 Sonnet will also answer more questions, meaning there will be fewer instances where it declines to respond. That’s because it’s better able to make a distinction between benign and harmful prompts, Anthropic said.

Coding models and more cash on the way

In addition to the reasoning model, Anthropic debuted a new model called Claude Code, which is available as a research preview now and is more specifically focused on coding tasks.

In a demonstration, the company showed how Claude Code is able to analyze a development project via a single prompt, such as “explain this project structure.” It also enables developers to modify a codebase simply by entering a plain English prompt that explains how they want it to alter the code. After making its changes, it will describe the edits it has made, and then test it for errors or push the update to a GitHub repository.

The company said Claude Code is available for testing for a limited number of users, and is offering access on a first-come, first-served basis, so developers who want to check it out should not delay.

The new models announced today represent an impressive breakthrough for Anthropic, and there could soon be many more developments on the way, for the company is said to be in advanced talks over a $3.5 billion funding round, according to a separate report today from the Wall Street Journal.

That amount is significantly higher than the initial $2 billion it had first set out to raise. It would increase the startup’s valuation to about $61.5 billion, the Journal said, citing two anonymous sources familiar with the talks. It’s said that Lightspeed Venture Partners will lead the round, with General Catalyst and various others also participating.

Image: SiliconANGLE/Meta AI