Claude 3.7 Sonnet, the first AI with hybrid reasoning: choose whether to activate your advanced reasoning capabilities
In recent months we have attended a proliferation of reasoning, capable of dividing tasks into steps and executing a thought process that culminates in more refined responses. The last of these tools is Claude 3.7 Sonnetthe most advanced model of the developer Anthropic. However, this is a particularity: it is a Hybrid Reasoning Model.
This means that, unlike other reasoning, Claude 3.7 Sonnet allows the user to activate or not these advanced thinking capabilities. In this way, when you want the AI of a quick and simple response, you will only have to request it, and the same when you need deep reflections for complex tasks. «We believe that Reasoning should be an integrated capacity of avant -garde models Instead of a completely separate model, ”they explain from Anthropic.
In addition to this novelty, Anthropic has also presented Claude Codean agents coding tool that is in preliminary phase. With Claude Code, developers and developers can delegate complex engineering tasks to Claude directly from their terminal.
Introduction Claude 3.7 Sonnet: Our Most Intelligent Model To Date. IT's a Hybrid Reasoning Model, producing near-instant Respons Respidad, Step-by-Step Thinking.
One Model, Two Ways to Think.
We're Also Releaseing An agentic coding Tool: Claude Code. pic.twitter.com/jt7qmfwuc
– Anthropic (@anthropicai) February 24, 2025
How Claude 3.7 Sonnet works, the hybrid reasoning AI
Claude 3.7 Sonnet is the Evolution of Claude 3.5 Sonnet (Indeed, 3.6) have been skipped and it is the first reasoning AI developed by Anthropic. In addition, he presents improved capabilities In mathematics, physics, instruction or coding monitoring, among other areas.
Being a hybrid modelintegrates both the capacities of a common LLM and a reasoning. The user or user will only have to deploy the “Claude 3.7 Sonnet” button located in the AI text drawer itself and select the option “Normal” or “Extended”depending on whether or not you need your reasoning capabilities.
This characteristic The difference of other reasoning models such as O3 of OpenAIthat only allow to execute complex thinking processes and steps to respond to consultations. Which implies that the user or user has to select a different model according to the complexity of the task. In fact, OpenAI recently reported his Plans to delete the chatgpt model selector and develop a tool capable of applying the most appropriate model in each context, seeking to simplify and improve the user experience.
This It is also the goal of Anthropic. The Laboratory of AI created by former OpenAi employees wishes to move towards a model capable of deciding how much “think” or “reason” a task, eliminating the intermediate step that forces users and users selecting the “normal” or “extended mode or” extend ».
Optimized for the real world
In the chat panel itself, Claude 3.7 Sonnet will show the internal reasoning process that performs until you reach the final answer And it will mark the time it has taken in reaching its conclusion. Of course, from Anthropic they point out that it will not always reveal all their “thoughts”, since some may be censored for security reasons.
The developer has optimized the ways of thinking of this artificial intelligence to perform real world tasks that reflect how companies use this technology to improve their productivity, such as coding problems or agency tasks.
Likewise, Claude 3.7 Sonnet has improved its ability to identify harmful applications and differentiate them from those that are not. Version 3.7 has reduced the rate of unnecessary rejections by 45% compared to its predecessor 3.5.
Regarding performance of this new model, in the Swe-Bench test (coding tasks) revealed a 62.3% precisionwhile OPENAI's O3-mini obtained 49.3%. And, in the Tau-Bench testwhich measures the ability of an AI model to interact with simulated users and external APIS, Claude 3.7 Sonnet achieved a 81.2%compared to OPENAI O1, which obtained 73.5%.
A Pokemon Gymnasium Training
As a curious detail, it is worth mentioning that Anthropic not only uses these official tests to Test your new AIbut also resorted to other ways such as Play red pokemon video game of Game Boy.
To do this, as explained, «we equip Claude with basic memory, pixels entry on the screen and function calls to press buttons and navigate the screen, which allowed him to play Pokémon continuously beyond his limits of usual context, keeping the game along tens of thousands of interactions ».
Claude 3.7 Sonnet managed to defeat three gym leaders and win their medals. This has been the best result obtained by a model of the Claude Sonnet family, whose first Claude 3.0 version Sonnet did not even get out of the house in Paleta town.
«Pokémon is a fun way to appreciate Claude 3.7 Sonnet, but we hope that these capabilities have an impact on the real world far beyond games. The model capacity to maintain concentration and achieve open goals It will help developers to create a wide range of the latest -generation AI agents, ”says the developer.
Reasoning capacity is only available in premium plans
Anthropic has given Claude 3.7 Sonnet to all its users and users, however not all plans give access to its full version. The ability to reasoning of the “extended” mode will only be available for those who have contracted A PAYMENT PLAN (PRO, Team or Enterprise). For its part, the free plan will offer the improved capabilities of Claude 3.7 Sonnet as Model LLM, but without the reasoning function.
It is also possible to use Claude 3.7 Sonnet and its reasoning capabilities in Anthropic Api, Amazon Bedrock and Verex Ai from Google Cloud.
When using this model Through the API«Users can also control the budget to think: They can indicate to Claude that you think for no more than n tokens, for any value of N until its output limit of 128k tokens. This allows them to balance the speed (and cost) for the quality of the response ».
Claude Code, an active agent for coding tasks
The other novelty presented by Anthropic is Claude Code, Your first active agent specialized in coding tasks. This model can “search and read code, edit files, write and execute tests, confirm and send code to GITHUB and use command line tools, keeping it informed in each step.”
At the moment, Claude Code is in a PREVIOUS VIEW OF LIMITED INVESTIGATIONbut their abilities have already revealed great results. From the developer they explain that this model managed to complete tasks in a single pass thatnormally, They would take more than 45 minutes of manual work.
Anthropic has reported that he plans Implement continuous improvements relying on the experience of use. Specifically they contemplate: “Improve the reliability of tools calls, add support for prolonged execution commands, improve representation in the application and expand Claude's own understanding about their abilities.”
It is already possible to request access to the preliminary version of Claude Code by pointing to your waiting list.
Photo: Anthropic
