Anthropic Unveils Claude 3.7 Sonnet: The World’s First Hybrid Reasoning Model

Anthropic has launched Claude 3.7 Sonnet, the first hybrid reasoning model that combines traditional large‑language‑model generation with a dedicated reasoning phase. The announcement notes that OpenAI’s upcoming GPT‑5 will be a unified model, while Claude 3.7 already offers fast responses and deeper reasoning.

Extended Thinking Mode lets the model spend configurable time “thinking” before answering. Users can see the reasoning trace, and developers can control the depth by specifying a token budget (up to 128 K tokens) through the API.

import Anthropic from '@anthropic-ai/sdk';
const client = new Anthropic();
const response = await client.messages.create({
  model: "claude-3-7-sonnet-20250219",
  max_tokens: 20000,
  thinking: {
    type: "enabled",
    budget_tokens: 16000
  },
  messages: [{ role: "user", content: "Are there an infinite number of prime numbers such that n mod 4 == 3?" }]
});
console.log(response);

The API response contains both a thinking block and a text block, e.g.:

{
  "content": [
    { "type": "thinking", "thinking": "To approach this, let's think about what we know about prime numbers..." },
    { "type": "text", "text": "Yes, there are infinitely many prime numbers such that..." }
  ]
}

Larger Output Capacity expands the maximum output to 128 K tokens—a 15× increase over previous versions—enabling complex code generation, detailed design documents, long‑form writing, and large‑scale data analysis without hitting token limits.

Improved Coding Abilities are highlighted as a major breakthrough. The model excels at planning and solving intricate coding tasks, handling large codebases, building full‑stack web applications, and producing production‑ready code. Early tests by companies such as Cursor, Cognition, Vercel, and Replit show superior performance on real‑world coding workloads.

Reduced Unnecessary Refusals cut the rate of unwarranted rejections by 45 % compared with prior models, making the system safer yet more helpful.

Claude Code is a new command‑line tool designed for “agentic coding.” It lets developers assign major engineering tasks from the terminal, supporting code search, file editing, test writing/running, GitHub commits, and step‑by‑step execution. Initial tests report an 80 % reduction in development time for tasks that previously required 45 minutes of manual work.

Performance benchmarks show top‑tier results: Claude 3.7 Sonnet ranks at the forefront of the SWE‑bench Verified suite for software‑engineering problems and achieves leading scores on the TAU‑bench framework for multi‑step reasoning and tool use.

Access is provided through the Claude web interface, iOS and Android apps, and the public API. Users simply select “Claude 3.7 Sonnet” from the model dropdown. The extended thinking mode is available only on paid plans (Professional, Team, Enterprise).

Pricing mirrors earlier models: free accounts receive the base model without extended thinking; Professional plans cost $20 / month, Team plans $30 / month. Token pricing is $3 per million input tokens and $15 per million output tokens, with a 50 % discount for batch processing.

In summary, the deeper reasoning capability reduces hallucinations and yields more factual, trustworthy answers for researchers, while ordinary users benefit from more reliable, natural‑sounding conversations and longer context windows.