Anthropic has released a new AI model, Claude 3.7 Sonnet, which is the only one so far that can provide both real-time answers and more considered, “thoughtful” answers to questions. Users can choose whether to activate the AI model’s “reasoning,” which prompts Claude 3.7 Sonnet to “think” for short or long periods.
Anthropic has released a new AI model, Claude 3.7 Sonnet, which is the only one so far that can provide both real-time answers and more considered, “thoughtful” answers to questions. Users can choose whether to activate the AI model’s “reasoning,” which prompts Claude 3.7 Sonnet to “think” for short or long periods.
Claude 3.7 Sonnet became available on Monday to all users and developers. But access to the model reasoning features will be available to those on the premium plan. Free users will get the standard version of Claude 3.7 Sonnet, which Anthropic claims is superior to the previous Claude 3.5 Sonnet. If you noticed, the company skipped one digit — 6, meaning that after 3.5 comes 3.7.
Claude 3.7 Sonnet with reasoning is more expensive than other “thinking” models, but it’s important to remember that it’s a hybrid model. So Claude 3.7 Sonnet costs $3 per 1 million input tokens (meaning that for $3 you can feed Claude about 750,000 words—more words than in the entire Lord of the Rings series) and $15 per 1 million output tokens. This makes it more expensive than OpenAI’s o3-mini ($1.10 per 1 million input tokens/$4.40 per 1 million output tokens) and DeepSeek’s R1 ($55 cents per 1 million input tokens/$2.19 per 1 million output tokens).
Reasoning models, such as Google's o3-mini, R1, Gemini 2.0 Flash Thinking, and xAI's Grok 3 (Think), use more time and processing power before answering a question. The models break down tasks into smaller steps, which typically improves the accuracy of the final answer. Reasoning models do not necessarily think or reason like humans, but their process is modeled on the principle of deduction.
Ultimately, Anthropic would like Claude to figure out how long it should “think” about questions on its own without requiring users to select controls in advance, Anthropic’s head of product and research, Diana Penn, told TechCrunch.
Anthropic says it allows Claude 3.7 Sonnet to show the internal planning phase through a “visible notebook.” Users will see the full thought process of the AI model for most prompts, but some parts can be redacted for trust and security purposes.
The company claims to have optimized Claude's thinking modes for real-world tasks, such as complex coding problems or agent-based tasks. Developers using Anthropic's API can control the "budget" for thinking, the speed of trade, and the cost of response quality.
In one test, SWE-Bench, to accurately assess the ability of AI models to solve real-world software problems, Claude 3.7 Sonnet scored 62.3% accuracy, compared to OpenAI’s o3-mini model, which scored 49.3%. In another test, TAU-Bench, to measure the ability of an AI model to interact with simulated users and external APIs in retail, Claude 3.7 Sonnet scored 81.2%, compared to OpenAI’s o1 model, which scored 73.5%.
Anthropic also says that the Claude 3.7 Sonnet will refuse to answer questions less often than its previous models, claiming that the model is able to make finer distinctions between malicious and benign prompts. The company says this has reduced unnecessary refusals by 45% compared to the Claude 3.5 Sonnet.
In addition to Claude 3.7 Sonnet, Anthropic is also releasing an agent coding tool called Claude Code. Launched as a research release, this tool allows developers to run specific tasks through Claude directly from their terminal.
As an Anthropic representative told TechCrunch, Claude Code will initially be available to a limited number of users on a first-come, first-served basis.
