your123 - stock.adobe.com
Anthropic intros hybrid reasoning model: Claude 3.7 Sonnet
The model can perform both generic LLM and reasoning tasks. It is optimized for business use. The vendor also introduced an agent for generating code.
Anthropic on Monday introduced a new hybrid reasoning model and coding agent.
Claude 3.7 Sonnet can produce near instant or step-by-step thinking responses, according to the independent generative AI vendor.
It is available on all Claude plans, through the Anthropic API and on Amazon Bedrock and Google Cloud Vertex AI.
Compared to other reasoning models, Claude 3.7 Sonnet is both a standard LLM that performs text summarization and generic tasks and a reasoning model, Anthropic said. Users can pick how the model can respond. In the standard, or regular large language model (LLM) mode, it's an upgraded version of the previous generation 3.5 Sonnet.
However, Claude 3.7 Sonnet self-reflects in an extended thinking setting, improving its performance in math, physics, coding and other tasks, API users can also control how long the model can think for and pay for as many tokens as they need, up to 128K tokens. The model is optimized for business applications such as coding.
Anthropic also introduced a new coding agent, Claude Code. The vendor said Claude Code completed tasks that would normally take 45 minutes or more to complete quickly. While the agent is still in its early iteration, Anthropic plans to improve it in the coming weeks.
The release of Claude 3.7 Sonnet and Claude Code reflects a generative AI market that is currently focused on two big themes: reasoning models and agents.
A reasoning model
Reasoning models have made big waves in the past few weeks after AI Chinese startup DeepSeek's release of its R1 model. Before DeepSeek-R1, OpenAI and Google also introduced reasoning models.
"The market is moving toward building these more capable reasoning models," said Gartner analyst Arun Chandrasekaran. "The overall movement toward reasoning is not surprising at all. We will see a lot more of this in 2025."
Anthropic's approach to AI reasoning differs from that of other providers. Most AI providers have created a separate reasoning model instead of combining reasoning with the regular LLM.
There is an advantage to having an LLM that can both reason and be a standard LLM, said Rowan Curran, an analyst with Forrester Research.
"[If] you're trying to make the model as generalized as possible, it may not be as good as the specific task that you want to be good at," Curran said. A hybrid model can perform well in both reasoning and summarization.
Moreover, Curran added, the model can be trained for real-world business tasks rather than just math and science.
"Seeing how enterprises are starting to use these as ways to plan and resolve toward goals, rather than doing math specifically, is a good start and interesting way to produce potentially good results," he said.
Agentic AI
Besides focusing on reasoning, Anthropic's Claude Code follows the much-hyped agentic AI trend.
Since 2024, agentic AI spun off Salesforce Agentforce, Microsoft Agents and Google Agentspace, among other agentic platforms.
On Monday, Salesforce and Google revealed that Salesforce customers can build Agentforce agents using Gemini and deploy Salesforce on Google Cloud.
AI search vendor Perplexity also said on Monday it will soon release a browser for agentic search called Comet.
Claude Code is not only a way for Anthropic to bring an agentic AI system to market but also lets the vendor to play to its strength.
"Anthropic has been very good at coding-oriented tasks," Chandrasekaran said. "I would imagine that they want to go deeper into that domain, which is trying to automate software development lifecycle workflows."
Curran said a coding agent is also beneficial in a in which code generation will be one of the key agentic capabilities for generating smaller pieces of code, to perform small tasks or API calls.
"Having that as part of their offering is going to be a necessary part of being able to provide these types of capabilities," he said.
Meanwhile, Chandrasekaran said a challenge for Anthropic with agentic AI is gaining enough traction to become successful.
"I'm very interested to see how Anthropic can build the right go-to-market to make the agent aspect successful, as well as the application aspect of it successful," he said.
Esther Shittu is an Informa TechTarget news writer and podcast host covering artificial intelligence software and systems.
