The startup Anthropic, which is based in San Francisco and operates in the space of the artificial intelligence industry, presented a new frontier AI model called Claude 3.7 Sonnet.
The mentioned startup stated that its new virtual product was elaborated to think about issues as long as consumers needed. Anthropic described Claude 3.7 Sonnet as the industry’s first hybrid artificial intelligence reasoning model. In this case, the startup offers consumers a single machine intelligence model that can provide both real-time answers and more considered, thought-out answers to questions. Users can choose whether to activate Claude 3.7 Sonnet’s reasoning abilities, which prompt the virtual product to think for a short or long period.
The new artificial intelligence model reflects the startup’s broader efforts aimed at simplifying the process of consumer interaction with its AI products. It is worth noting that currently most chatbots have a daunting model picker. As part of the relevant practice, users should choose between several options that differ from each other in terms of parameters such as cost and capabilities. Labs like Anthropic are focused on achieving positive outcomes as part of their efforts to ensure that consumers don’t have to make the mentioned choice. In this case, the goal is to make that all the work is done by one artificial intelligence model.
Anthropic stated that Claude 3.7 Sonnet is available to all users and developers. At the same time, the startup noted that access to the reasoning functions of the model will be provided only to those who pay for the premium tariff plan of the Claude chatbot. Users of the free version of the mentioned chatbot will interact with the standard configuration of the new virtual product from Anthropic, which is non-reasoning. The startup claims that Claude 3.7 Sonnet outperforms its previous frontier AI model called Claude 3.5 Sonnet.
The cost of the new virtual product from Anthropic is $3 per million input tokens. This means that for the specified amount, the user will be able to enter approximately 750,000 words into Claude, which is more than in the entire Lord of the Rings series. The cost of one million output tokens in this case is $15.
It is worth noting that the new virtual product from Anthropic is more expensive than the o3-mini from OpenAI. The ChatGPT developer offers users access to its mentioned artificial intelligence system at a price of $1.1 per million input tokens and $4.4 per million output tokens. DeepSeek’s R1 is even cheaper. In this case, users pay 55 cents per million input tokens and $2.19 per million output tokens. At the same time, it is worth paying special attention to the fact that the two mentioned artificial intelligence models are strictly reasoning. Claude 3.7 Sonnet has been described as a hybrid AI model.
The new virtual product from Anthropic is the first artificial intelligence model that can reason. Many machine intelligence labs have paid attention to the relevant technique, as traditional ways to improve the performance of digital intelligence taper off.
Such reasoning artificial intelligence models as o3-mini, R1, Google’s Gemini 2.0 Flash Thinking, and xAI’s Grok 3 (Think) use more time and computing power before answering questions. These models break problems down into smaller steps. The appropriate algorithm of actions contributes to a higher accuracy of the final answer.
It is worth noting that artificial intelligence reasoning models do not necessarily perform cognitive operations such as thinking or reasoning in the paradigm of the human approach to the relevant activity. At the same time, the process of functioning of these AI systems is modeled after deduction.
The ultimate goal of Anthropic is for the artificial intelligence model to independently figure out how long it should think about issues, without requiring users to select controls in advance. Startup’s product and research lead, Dianne Penn, told about this during a conversation with media representatives.
In a post published on the Anthropic blog, it was noted that just as humans do not have two separate brains that answer questions that can be answered immediately and those that require thought, the startup regards reasoning as one of the capabilities needed for frontier artificial intelligence models. It was also emphasized in the relevant context that the mentioned feature should be smoothly integrated with other capabilities, rather than being something that is provided in a separate machine intelligence model.
The startup stated that it allows Claude 3.7 Sonnet to display the internal planning stage through a visible scratch pad. Dianne Penn noted that users will see the entire Claude’s thinking process for most prompts, but some portions may be redacted for trust and safety purposes. Anthropic stated that it has optimized the mentioned virtual product’s thinking modes for real-world tasks such as difficult coding problems or agentic tasks. The developers tapping startup’s API can control what can be called the budget for thinking, the speed of trading, and the cost for quality of answer.
On one test to measure real-word coding tasks, SWE-Bench, Claude 3.7 Sonnet demonstrated an accuracy of 62.3%. It is worth noting that a similar indicator of the o3-mini from OpenAI was recorded at the 49.3% mark. In another test to measure the artificial intelligence model’s ability to interact with simulated consumers and external APIs in a retail setting, TAU Bench, Claude 3.7 Sonnet scored 81.2%. The result of OpenAI’s o1 in this case was 73.5%.
Anthropic stated that the Claude 3.7 Sonnet will refuse to answer questions less often than its previous models. The startup claims that its new virtual product is able to make more nuanced distinctions between harmful and benign prompts. Anthropic stated that the number of unnecessary refusals decreased by 45% compared to the figure demonstrated by Claude 3.5 Sonnet. It is worth noting that some other artificial intelligence labs are currently rethinking their approach to restricing the answers of their chatbots.
Anthropic also releases an agentic coding tool called Claude Code. It is worth noting that this tool was launched as a research preview. The mentioned virtual product helps developers to run specific tasks through Claude directly from their terminal.
In the demo, employees of Anthropic showed how Claude Code can analyze a coding project with a simple command. A developer can modify a codebase using plain English on the command line. Claude Code will describe its edits as changes are made and even test the project for errors or push it to the GitHub repository.
An Anthropic spokesperson stated in a media comment that the mentioned tool will be available to a limited number of users on a principle that can be formulated as first come, first serve.
Claude 3.7 Sonnet’s debut comes at a time when artificial intelligence labs are launching new AI models at an incredible rate. Anthropic adheres to a more methodical approach focused on safery. Currently, the startup has probably decided to accelerate, looking to lead the pack. It is worth noting that in the area of machine intelligence, hybrid models in the foreseeable future may cease to be a rare or unique product with an non-minimal probability. If the Claude 3.7 Sonnet example proves successful, it is likely that other AI players will elaborate hybrid models. OpenAI is already working on a corresponding project. The startup’s chief executive officer, Sam Altman, said that the appropriate product from the ChatGPT developer will arrive in months.
It is worth noting that against the background of the active development and scaling of artificial intelligence as an advanced technology of modernity, the issue of cybersecurity is growing in urgency. Scammers also have access to AI tools, which is why their activities have become more sophisticated. To counteract the corresponding threat in the virtual space, personal awareness of users is important. For example, an Internet search query such as how to know if my camera is hacked will allow anyone to get information about signs of unauthorized access to the device. Digital literacy is an effective tool for countering cybercrime. At the same time, relevant knowledge should be updated periodically, as cybercriminals seek to use the most advanced technologies in their activities. Moreover, artificial intelligence-based tools are currently being actively developed to detect and eliminate security threats in the virtual space.
