Anthropic's new model is about to be released! The "sliding ratio" helps companies reduce costs and increase efficiency, with programming capabilities surpassing o3-mini

Author of this article: Fang Jiayao

Source: Hard AI

On February 13th, Eastern Time, The Information reported that the artificial intelligence startup Anthropic plans to launch an innovative hybrid AI model in the coming weeks.

The uniqueness of this model lies in its ability to dynamically switch between deep reasoning and quick responses, providing developers with a multifunctional tool. According to a person who has used the model, Anthropic's new model can mobilize more computational resources for in-depth reasoning when dealing with complex problems, while being able to process simple tasks quickly like traditional large language models without additional computation.

Most importantly, it offers a novel "sliding scale" feature, allowing developers to precisely control the amount of computational resources the model uses when processing queries, thereby adjusting the depth of reasoning and the model's "thinking" time. For example, setting the slider to "0" will make the model operate as a regular non-reasoning AI, similar to OpenAI's GPT-4o.

This enhances cost-effectiveness for developers, enabling them to balance cost, speed, and performance according to specific needs. In contrast, OpenAI's design choices regarding reasoning capabilities are relatively simple. The company currently offers three fixed settings: "low," "medium," and "high." Some developers have reported that the actual number of tokens processed by the predictive model at different levels remains uncertain, making it difficult for them to estimate the cost of a single query.

From a product design perspective, OpenAI seems to focus more on ordinary consumers and individual professional users, hence adopting a simple and understandable "low-medium-high" naming convention, while Anthropic is more focused on the enterprise market, with its refined control feature allowing developers to adjust the AI model's computational resources via a slider, enabling businesses to find the best balance between cost, speed, and performance.

Users on social media platform X have expressed:

"Given that Claude has already caused a stir, this could change the game."

"Get ready to be blown away by another AI model! Anthropic is about to launch a hybrid genius that can think deeply and respond faster than you can say 'innovation.'"

Does Anthropic's coding ability surpass OpenAI's o3-mini?

Last autumn, OpenAI was the first to release a model with reasoning capabilities, followed by companies such as Google, Alibaba, and GaoFei Capital Management, which spurred competition in reasoning models In this competition, Anthropic has remained low-key, choosing a different technological path. Now, it has finally unveiled the mystery, planning to achieve ambitious revenue goals with this new model. According to a report by The Information, Anthropic aims to surpass OpenAI and become the preferred AI supplier for application developers.

According to OpenAI CEO Sam Altman's latest disclosure on social media platform X, OpenAI will launch GPT-4.5, which will be a traditional non-reasoning model. Next, OpenAI plans to integrate the GPT model with the "o" reasoning model into a single AI system, a move that seems to draw inspiration from Anthropic's research and development path.

It is worth mentioning that Anthropic's new model has made significant progress in programming. Generative AI has become one of the most powerful application areas in enterprise applications, with programming at the forefront. A person who has used the model stated that when the model is allowed to "think" for the longest time, it outperformed the most advanced OpenAI reasoning model currently available to customers (i.e., the o3-mini model set to "high" reasoning level) in certain programming benchmark tests.

Unlike OpenAI's reasoning model, which is better at handling academic programming problems, Anthropic's model is more suited for the daily programming tasks of enterprise engineers. Particularly when dealing with complex codebases, Anthropic's model can better understand intricate codebases composed of thousands of files and can generate complete lines of code that run successfully on the first attempt.

This feature is especially noteworthy, as AI models often tend to "cut corners" when faced with complex programming problems, providing comments instead of actual code.

Currently, the pricing issue regarding Anthropic's new model remains unclear. The industry is generally concerned about whether it can be more cost-effective than OpenAI's recently released o3-mini reasoning model, which has been considered to have a high cost-performance ratio, attracting a large number of developers.

With the launch of low-cost AI products like DeepSeek and Google's latest Gemini model, whether the entire market will trend towards lower prices remains a hot topic of concern in the industry. Currently, the AI model market may be divided into two camps: one is cutting-edge development, where models that continuously improve reasoning or programming capabilities still have pricing advantages. The other is "good enough" models, which may drive each other's prices down, leading to fierce competition