🤖 AI Agents Weekly: Claude Haiku 4.5, Deep Agents, SWE-grep, nanochat, Agent Skills, Veo 3.1 Fast, n8n AI Workflow Builder

Claude Haiku 4.5, Deep Agents, SWE-grep, nanochat, Agent Skills, Veo 3.1 Fast, n8n AI Workflow Builder

In today’s issue:

• Anthropic introduces Claude Haiku 4.5

• Deep Agents: The future of AI Agents

• Demystifying RL in Agentic Reasoning

• Cognition introduces SWE-grep for faster context retrieval

• Andrej Karpathy releases nanochat

• Anthropic introduced Agent Skills

• n8n introduced AI Workflow Builder

• OpenAI released gpt-5-search-api

• Google DeepMind and Yale launched Cell2Sentence-Scale 27B

• Google introduced Veo 3.1 and Veo 3.1 Fast

• Top AI dev news, tools, and papers.

---

---

Top Stories

Claude Haiku 4.5

Anthropic’s new small model, Claude Haiku 4.5, delivers near–frontier performance at a fraction of the cost and latency. It matches or beats Claude Sonnet 4 on real-world coding and computer-use benchmarks, making it ideal for responsive AI tasks such as live chat, coding agents, and customer-service automation.

• Speed and efficiency – Haiku 4.5 runs 2× faster and costs ⅓ as much as Sonnet 4, while achieving comparable SWE-bench Verified scores and even outperforming it on computer-use tasks.

• Ideal for multi-agent workflows – Anthropic highlights pairing Sonnet 4.5 (for planning) with teams of Haiku 4.5s (for execution) to parallelize subtasks efficiently.

• Safety-first design – In alignment testing, Haiku 4.5 showed the lowest rate of misaligned behaviors among all Claude models, earning the AI Safety Level 2 (ASL-2) classification, less restrictive than Sonnet 4.5’s ASL-3.

• Broad deployment – Available via the Claude API, Claude Code, Amazon Bedrock, and Google Vertex AI, it’s positioned as a drop-in replacement for Haiku 3.5 or Sonnet 4 for developers seeking cost-efficient, real-time reasoning.

• Benchmarks – On SWE-bench Verified, Terminal-Bench, τ²-Bench, AIME, OSWorld, and MMMLU, Haiku 4.5 maintains strong accuracy using 128K thinking budgets with tool use enabled.

Blog