AI Agents Weekly: 4o Image Gen, Gemini 2.5 Pro, Agents SDK Guide, AgentRxiv

4o Image Gen, Gemini 2.5 Pro, Agents SDK Guide, AgentRxiv

In today’s issue:

• OpenAI announces 4o image generations

• Google DeepMind introduced Gemini 2.5 Pro

• New OpenAI Agents SDK Guide

• Why Qodo Chose LangGraph

• DeepSeek releases DeepSeek-V3-0324

• Ai2 Paper Finder agent

• Qwen2.5-Omni is a new multimodal flagship model from Qwen

• LLM agents to autonomously generate and share research papers

• Top AI dev news and much more.

---

---

Top Stories

OpenAI’s 4o Image Generations

OpenAI has unveiled 4o Image Generation, a new multimodal extension to GPT‑4o that merges advanced text-based reasoning with powerful image creation. This release emphasizes usefulness, from precise text rendering on graphics to maintaining consistency in multi-turn prompts, while providing photorealistic, context-aware outputs. Below are the core highlights:

• Built-In Multimodality – The model unifies text and image generation natively, letting users refine and iterate on images within ChatGPT and maintain visual consistency across multiple turns.

• High-Fidelity Text Rendering – Unlike prior image generators that often distort lettering, 4o Image Generation can accurately produce symbols, signs, and typed text, aiming to support creative workflows like custom menu designs, diagrams, or signage.

• Robust Instruction Following – GPT‑4o can now handle prompts with numerous specific details (10–20 distinct objects or instructions) and preserve relationships among them. It keeps characters, objects, and textual elements consistent over extended tasks.

• Multi-Turn Enhancements – Users can refine images by referencing prior outputs in the same conversation (e.g., “add a detective hat” or “give the cat a monocle”), ensuring cohesive transformations without losing context.

• Practical Use Cases – Beyond novelty images, 4o Image Generation is tailored for diagrams, infographics, and product designs. For example, generating data-driven visuals or accurately depicting brand assets—tasks that historically challenged generative models.

• Limitations and Ongoing Improvements – The system occasionally struggles with cropping longer images, precise edits on small text, non-Latin scripts, and complex or “dense” textual content. Editing partial regions may inadvertently affect other parts of the image.

• Safety and Provenance – All generated images carry C2PA metadata to identify them as AI-generated. The model blocks disallowed requests (e.g., harmful or sexual content) and uses internal verification to mitigate misuse. Future efforts aim to refine face editing and overall reliability.

• Availability – Rolling out initially to Plus, Pro, Team, and Free ChatGPT users, 4o Image Generation will soon expand to Enterprise, Education, Sora, and an API for developers. Images typically take up to a minute to render, given the model’s higher detail.

Blog