🧠 AI Agents Weekly: Claude 3.7 Sonnet, GPT-4.5, VisionAgent, Phi-4, Thinking QwQ

Claude 3.7 Sonnet, GPT-4.5, VisionAgent, Phi-4, Thinking QwQ

In today’s issue:

• Anthropic announces Claude 3.7 Sonnet

• OpenAI launched GPT-4.5

• VisionAgent is an open-source Python library using agents for vision tasks

• DeepSeek Open-source Week: FlashMLA, DeepEP, 3FS,…

• Microsoft announced Phi-4

• Qwen launches Thinking (QwQ)

• Google’s new method for improving planning and reasoning with agents

• Top AI dev news and much more.

---

---

Top Stories

Claude 3.7 Sonnet

Anthropic has unveiled Claude 3.7 Sonnet, its most advanced AI model yet, alongside Claude Code, a new agentic coding tool.

Key highlights:

• Hybrid reasoning model – Claude 3.7 Sonnet is the first AI model that dynamically switches between rapid responses and in-depth, step-by-step reasoning, controlled via API settings.

• State-of-the-art coding performance – Benchmarks like SWE-bench Verified and TAU-bench show Claude 3.7 Sonnet leading in real-world coding tasks, outperforming previous models.

• Extended thinking mode – Available on Pro, Team, and Enterprise plans, this feature allows deeper self-reflection for better accuracy in math, physics, and instruction-following.

• Fine-tuned control over response depth – API users can limit the model’s "thinking budget" by setting a maximum number of reasoning tokens.

• Launch of Claude Code – A command-line coding assistant that automates test-driven development, debugging, and large-scale refactoring directly from the terminal.

• GitHub integration – Now available across all Claude plans, enabling seamless repository connections for debugging, documentation, and full-stack development.

• Safer & more reliable – Reduces unnecessary refusals by 45% and includes enhanced defenses against prompt injection attacks.

Blog | Claude Code Overview | Extended Reasoning Report