Humanity is releasing a new frontier AI model called Claude 3.7 Sonnet, designed to “think” about questions as long as users want.
Humanity calls Claude 3.7 Sonnet the industry’s first “hybrid AI inference model.” This is because it is a single model that can give both real-time answers and “conceived” answers to more considered questions. Users can choose whether to activate the “inference” ability of the AI model. This encourages Claude 3.7 Sonnet to “think” for a short or long period of time.
This model represents humanity’s broad commitment to simplifying the user experience around AI products. Most AI chatbots today have a difficult model picker that allows users to choose from several different options with different costs and capabilities. Labs like humanity don’t need to think about it. Ideally, one model will do all the work.
Claude 3.7 Sonnet is rolling out to all users and developers on Monday, Humanity said, but only those who pay for humanity’s premium Claude Chatbot plan will have access to the model’s inference capabilities. Free Claude users get the standard, irrational version of Claude 3.7 Sonnet. This claims that humanity outweighs the previous frontier AI model, the Claude 3.5 Sonnet. (Yes, the company skipped the number.)
Claude 3.7 Sonnet costs $3 for an input token per million dollar (i.e. about 750,000 words, which allow you to type more words than the entire “Lord of the Rings” series. This will allow Openai’s O3 -It’s more expensive than Mini ($1.10 per million input token/$4.40 per million output token) and Deepseek’s R1 (55 cents per million input token/$2.19 per million output token) , care, O3-MINI and R1 are strictly inference models, not hybrids like the Claude 3.7 Sonnet.
Claude 3.7 Sonnet is the first AI model for humanity that can “infer.” This is a technique that many AI labs have seen as a traditional way to improve AI performance taper.
Inference models such as O3-Mini, R1, Google’s Gemini 2.0 Flash Thinking, Xai’s Grok 3 use more time and computing power before answering the question. The model tends to break down the problem into smaller steps and improve the accuracy of the final answer. Inference models do not necessarily think or reason like humans, but the process is modeled after deduction.
Ultimately, humanity has led human products and research lead Dianne Penn to “think” about questions on its own, without the need for users to select controls beforehand, without the user having to select controls. I hope to understand.
“As humans don’t have two separate brains for questions that require thought to answer quickly,” says Humanity in a blog post shared with TechCrunch. I’m writing this way. Make sure it integrates smoothly with other features rather than being offered by another model. ”
Humanity says that the Claude 3.7 sonnet is allowing the internal planning stage to be displayed via a “visible scratchpad.” Penn said TechCrunch users will see the full thought process of Claude at most prompts, but some parts may be edited for reliability and safety.
Humanity says that it optimized Claude’s thinking mode for real tasks, such as difficult coding problems and agent tasks. Developers who tap on Anthropic’s API can control “budget” for thought, transaction speed and cost for answer quality.
In one test to measure Real Word coding tasks, the SWE bench, Claude 3.7 Sonnet, was 62.3% accurate compared to Openai’s O3-Mini model, which earned 49.3%. In another test measuring the ability of AI models to interact with simulated users and external APIs in retail settings, Tau-bench, Claude 3.7 Sonnet is 81.2% compared to Openai’s O1 model, which earned 73.5%. I’ve won.
Humanity also refuses to answer questions more frequently than previous models, claiming that the model can create a more subtle distinction between harmful and benign prompts. Humanity says it has reduced unnecessary rejections by 45% compared to the Claude 3.5 sonnet. This happens when some other AI labs are rethinking their approach to limiting AI chatbot answers.
In addition to Claude 3.7 Sonnet, Anthropic has also released an agent coding tool called Claude Code. When launched as an investigation preview, this tool allows developers to perform certain tasks from the terminal via Claude.
In the demonstration, human employees showed how Claude code analyzes coding projects with simple commands such as “Describe this project structure.” Using plain English on the command line, developers can change the codebase. Claude Code explains the edits when making changes, tests the error project, and pushes it to a GitHub repository.
Humanity spokesman is now available for a limited number of users on a “first cam, first serve” basis.
Humanity is releasing the Claude 3.7 sonnet when AI Labs ships new AI models at a breakneck pace. Historically, humanity has taken a more systematic, safety-focused approach. However, this time the company is about to lead the pack.
However, how much is the problem? Openai can be closer to the release of its own hybrid AI model. The company’s CEO, Sam Altman, said it will arrive on the “Moon.”
Source link
