🤖 AI Agents Weekly: GPT-5, Genie 3, gpt-oss, Cursor CLI, Opus 4.1, Efficient AI Agents

GPT-5, Genie 3, gpt-oss, Cursor CLI, Opus 4.1, Efficient AI Agents

In today’s issue:

• OpenAI announces GPT-5

• DeepMind introduces Genie 3

• OpenAI releases gpt-oss models

• Gemini CLI Deep Dive

• Cursor release Cursor CLI

• Anthropic announces Opus 4.1

• Groq has released Groq Code CLI

• New research on designing efficient AI Agents

• Top AI dev news, papers, tools, and much more.

---

---

Top Stories

GPT-5

OpenAI’s GPT-5 is the company’s most advanced model to date, unifying fast responses and deep reasoning within a single system that adapts dynamically to task complexity. A built-in router selects between a lightweight model for simple queries and a “GPT-5 thinking” mode for harder problems, with a mini fallback once usage limits are hit. GPT-5 significantly improves factual accuracy, instruction following, and style control while reducing hallucinations, sycophancy, and deceptive outputs.

It achieves state-of-the-art results in math (94.6% AIME 2025), coding (74.9% SWE-bench Verified), multimodal reasoning (84.2% MMMU), and health (46.2% HealthBench Hard), with GPT-5 Pro delivering even higher performance on complex, expert-level tasks.

Key advances include:

• Domain-specific improvements – Stronger coding capabilities (especially in complex front-end generation and repo-scale debugging), richer creative writing, and enhanced health advice with proactive question-asking and contextual adaptation.

• Evaluation gains – Outperforms GPT-4o and o3 across math, visual reasoning, agentic tool use, and economically valuable work, with GPT-5 Pro preferred by experts 67.8% of the time on challenging prompts.

• Reliability and safety – Hallucinations reduced by ~45% vs GPT-4o and ~80% vs o3; less than half the deception rate of o3; new “safe completions” training enables nuanced responses in dual-use domains; robust safeguards in high-risk bio/chemistry contexts.

• User experience enhancements – Reduced unnecessary agreement, improved custom instruction following, and new preset personalities for conversational style control.

• Efficiency – Delivers better results with 50–80% fewer reasoning tokens than o3, dynamically allocating computation.

Blog