The Batch: 885 | Claude Upgraded
Anthropic Upgrades Claude Sonnet to Version 4.5
原创 DeeplearningAI · 2025-10-14 12:46 · 北京
Anthropic has updated its medium-sized model Claude Sonnet, making it the first in the Claude series to reach version 4.5. In addition, the Claude Code intelligent coding tool has received important feature enhancements.
> Hi～新朋友，记得关注我们哟
---
Claude Sonnet 4.5 Overview
Key Highlights:
- Performance Boost: Significant improvements, plus the introduction of variable reasoning token budget.
- Inputs/Outputs:
- Accepts text and image inputs
- Depending on service tier, supports 200K – 1M tokens per input
- Outputs text up to 64K tokens
- Availability:
- Free via Claude.ai
- API access via Anthropic, Amazon Bedrock, Google Vertex
- Pricing: $3 per million input tokens, $15 per million output tokens
- Functionality:
- Variable reasoning budget
- Extended processing times (up to several hours)
- Sequential task execution (non-parallel)
- Knowledge Cutoff: January 2025
- Undisclosed: Model architecture, training data, training methods
---
Testing Results
- Text Ranking:
- With 32K reasoning token budget → #1 on LM Arena text leaderboard
- Without reasoning → #4 ranking
- SWE-bench Verified (coding challenge): 82% score — industry record; surpasses Claude Sonnet 4 (80.2%) and Claude Opus 4.1 (79.4%)
- OSWorld Benchmark (computer use): 61.4%, well above other models
- AIME 2025 (math):
- With Python tools → 100% accuracy
- Without tools → GPT-5 outperforms
- Visual Reasoning (GPQA-Diamond, MMMLU): Better than Claude Opus 4.1, slightly behind Google Gemini Pro 4.5 and OpenAI GPT-5
---
Claude Code Major Update
Claude Code has undergone a major upgrade, introducing new features aimed at developers and autonomous agents.
New Features
- Claude Agent SDK
- Built on Claude Code’s architecture, toolset, scheduling logic, and memory management system
- Offers foundational modules for web search, file management, code deployment
- Enables creation of autonomous agent applications
- Context Tracking
- Automatically generates summaries when message history nears capacity
- Removes unnecessary tool outputs to free space
- Memory Function
- `memory tool` API stores important external data (e.g., project status) for later retrieval
- Checkpoints
- Save rollback states to recover from errors
- IDE extensions (e.g., VSCode) replace command-line operations
---
Behind the News
Founded by former OpenAI staff, Anthropic positions itself as:
- Safer
- More human-centric
- More cautious than competitors
While keeping these values, it is increasingly focusing on coding and workplace productivity, targeting:
- Developers
- Enterprise users
As ChatGPT dominates consumer AI awareness, Anthropic’s developer-centric approach is a strategic positioning.
---
Key Insight
Combining Claude Sonnet 4.5 with the upgraded Claude Code reflects Anthropic’s aim for direct productivity gains — answering the recurring corporate question:
> “When will AI bring real productivity improvements to my team?”
AI-powered coding is currently one of the clearest pathways to these gains.
---
Our Take
The Claude Agent SDK is a milestone release — it could spark a surge in Claude-based agent innovations.
---
Additional Resources
---
Related Trend: AiToEarn
In the broader AI productivity ecosystem, platforms like AiToEarn官网 offer an open-source way to monetize AI-driven content and applications.
Features:
- Integrates AI content generation, cross-platform publishing, analytics, model ranking
- Supports publishing across platforms like Douyin, Kwai, WeChat, Bilibili, Xiaohongshu, Facebook, Instagram, LinkedIn, Threads, YouTube, Pinterest, and X (Twitter)
Parallel to Anthropic — AiToEarn also aims to make advanced AI capabilities accessible and impactful for creators, developers, and teams.
---
If you’d like, I can create a compact comparison table between Claude Sonnet 4.5 vs Claude Opus 4.1 vs GPT-5 to make this article even easier to digest. Would you like me to add that?
