原始文档 文章 Claude 2.1

Claude 2.1

文章 4 min read · 未标注

Release Date: November 21, 2023 Availability: API (via Console) and powering the claude.ai chat experience (free and Pro tiers).

Key Advancements

1. 200K Token Context Window

  • Capacity: Doubles the previous limit to 200,000 tokens, equivalent to roughly 150,000 words or over 500 pages of material.
  • Use Cases: Enables processing of entire codebases, financial statements (e.g., S-1s), or long literary works for summarization, Q&A, trend forecasting, and document comparison.
  • Note: Processing this context length is complex and may take a few minutes. Latency is expected to decrease over time. Access is reserved for Claude Pro users.

2. Reduced Hallucination & Improved Accuracy

  • Honesty: Achieves a 2x decrease in false statements (hallucinations) compared to Claude 2.0.
  • Testing Method: Evaluated on complex factual questions using a rubric that distinguishes incorrect claims from admissions of uncertainty. Claude 2.1 is more likely to demur when unsure.
  • Comprehension: Shows a 30% reduction in incorrect answers and a 3-4x lower rate of mistakenly concluding a document supports a particular claim, especially for long, complex documents (legal, financial, technical).

3. New Beta Feature: Tool Use

  • Functionality: Allows Claude to integrate with external tools, APIs, and data sources to perform actions on behalf of the user.
  • Examples of Use:
    • Using a calculator for complex numerical reasoning.
    • Translating natural language into structured API calls.
    • Answering questions via database or web search APIs.
    • Taking actions in software via private APIs.
    • Connecting to product datasets for recommendations.
  • Status: Early development; developer features and prompting guidelines are being built.

4. Enhanced Developer Experience

  • Workbench: A new playground-style interface for iterating on prompts, accessing model settings, and managing multiple prompts with saved revisions.
  • System Prompts: Allow users to provide custom instructions to set context, define personalities/roles, and structure responses for consistent, customizable outputs.
  • Code Generation: Developers can generate code snippets for their prompts to use directly in Anthropic's SDKs.

Pricing Update

  • Pricing has been updated to improve cost efficiency for customers across models.

来源

暂无来源