OpenAI has unveiled GPT-5.4, its latest frontier model designed for professional work, available in ChatGPT (as GPT-5.4 Thinking), the API, and Codex. A more powerful version, GPT-5.4 Pro, is also offered for users requiring maximum performance on complex tasks.

GPT-5.4 integrates recent advancements in reasoning, coding, and agentic workflows, incorporating the industry-leading coding capabilities of GPT-5.3-Codex while enhancing its performance across tools, software environments, spreadsheets, presentations, and documents. This results in more accurate and efficient task completion. Specifically, GPT-5.4 Thinking provides upfront thinking plans and improves deep web research, delivering higher-quality answers faster.

This model is the first general-purpose model with native, state-of-the-art computer-use capabilities, enabling agents to execute complex workflows across applications. It supports up to 1M tokens of context, allowing agents to plan, execute, and verify long-horizon tasks. The introduction of tool search significantly improves efficiency in large tool ecosystems by reducing token usage.

GPT-5.4 also boasts enhanced coding abilities, matching or outperforming GPT-5.3-Codex on SWE-Bench Pro and producing more aesthetic and functional results for frontend tasks. An experimental Codex skill, Playwright (Interactive), enables visual debugging and testing of web and Electron apps. Safety measures include an expanded cyber safety stack under the Preparedness Framework and improved Chain-of-Thought (CoT) monitorability.

👉 Read the full article on OpenAI

OpenAI researchers have published new findings on scattering amplitudes in quantum gravity, extending previous results for gluons to gravitons. This work demonstrates that a class of graviton interactions, long assumed to vanish, can in fact arise under specific kinematic conditions.

The paper, titled “Single-minus graviton tree amplitudes are nonzero,” authored by Alfredo Guevara, Alexandru Lupsasca, David Skinner, Andrew Strominger, and Kevin Weil, reveals that single-minus amplitudes (configurations where one particle has negative helicity and the rest have positive helicity) do not vanish when particle momenta satisfy a special alignment known as the half-collinear regime.

A significant aspect of this research is the contribution of GPT-5.2 Pro in discovering this new mathematical result. Provided with the earlier gluon paper as context, GPT-5.2 Pro was tasked with constructing the corresponding amplitudes in quantum gravity, a derivation that would have taken human authors considerable time. The model solved the problem using a surprising technique, the directed matrix-tree theorem, and produced an excellent preliminary draft of the paper.

This achievement represents a small step towards reconciling quantum mechanics with Einstein's theory of general relativity, illustrating how the infinite-dimensional “w-(1+∞)” symmetry, discovered by Roger Penrose, acts on gravitons. The project highlights how AI-assisted reasoning can contribute to theoretical research while upholding conventional standards of mathematical verification and scientific rigor.

👉 Read the full article on OpenAI

Luma, an AI video generation startup, has launched Luma Agents, creative AI agents powered by its new 'Unified Intelligence' models. These agents are designed to handle end-to-end creative work across text, image, video, and audio, proposing a new way of working for advertising agencies, marketing teams, design studios, and enterprises.

Luma Agents are built on the company's Uni-1 model, the first in its Unified Intelligence family. Uni-1 has been trained on audio, video, image, language, and spatial reasoning, enabling the agents to 'think in language and imagine and render in pixels or images.' This approach signifies a departure from the industry's prevailing method of chaining together separate models for language, vision, and generation, instead utilizing a single multimodal reasoning system for understanding and generation.

These agents coordinate multiple AI systems and work with other AI models, including Luma's Ray 3.14, Google Veo 3, ByteDance Seedream, and ElevenLabs' voice models. Luma states that the agents can evaluate and refine their own outputs through an iterative self-critique process. Demonstrations showed the agents generating various ad campaign ideas from a 200-word brief and an image, and localizing a $15 million ad campaign for different countries in just 40 hours.

Amit Jain, CEO and co-founder of Luma, emphasized that customers are not just buying a tool but are redoing how business is done, positioning Luma Agents as collaborators that allow creative teams to focus on taste, direction, and strategy rather than orchestrating tools. The platform has already been rolled out with global enterprise partners such as Publicis Groupe and Serviceplan Group.

👉 Read the full article on TechCrunch