🤖AI Agents Weekly: Gemini 2.5 Updates, Claude 4, II-Agent, Gemma 3n, MCIP, Veo 3

Gemini 2.5 Updates, Claude 4, II-Agent, Gemma 3n, MCIP, Veo 3

In today’s issue:

• Google I/O Gemini 2.5 Updates

• Anthropic announced Claude 4

• DeepMind’s new creative stack and Veo 3

• Microsoft Build updates and agent news

• OpenAI Response API updates

• Model Contextual Integrity Protocol

• Google announced Gemma 3n

• Running Claude’s coding agent programmatically

• GitHub has launched the Copilot coding agent

• II-Agent is a fully open-source, generalist AI assistant

• DeepMind is evolving Gemini 2.5 Pro into a world model

• Top AI devs news, research, and more

---

---

Top Stories

Google I/O Gemini Updates

Google’s latest 2.5-series release builds on the March launch with stronger reasoning, new multimodal abilities, and tighter security. Highlights:

• Performance leadership – Updated 2.5 Pro now tops WebDev Arena for coding and every LMArena leaderboard for overall quality. Its 1 M-token context window keeps state-of-the-art accuracy on long-context and video tasks, while LearnLM integration makes it the highest-rated model for pedagogy.

• Deep Think mode – An experimental setting for 2.5 Pro that tests multiple solution hypotheses before answering. Early scores: 2025 USAMO (math) SOTA, 84 % on MMMU (multimodal reasoning), and leadership on LiveCodeBench for competition-level coding.

• Faster, cheaper 2.5 Flash – The efficient sibling gains 20–30 % token savings yet improves on reasoning, multimodality, code, and long-context benchmarks. Preview is live in Google AI Studio, Vertex AI, and the Gemini app; GA in early June.

• New multimodal & agentic capabilities

  • Native audio output with tone/accent control, multi-speaker TTS, and emotion-aware dialogue in the Live API.

  • Project Mariner “computer use” skills (UI automation) exposed via Gemini API/Vertex AI for RPA partners this summer.

• Security and developer controls – Stronger defenses against indirect prompt injection; opt-in “thought summaries” expose model reasoning traces; thinking budgets let devs trade latency vs. quality; SDK now natively parses Model Context Protocol (MCP) tool definitions.

Blog