GPT-Image 2.5 Lands Everywhere, With Two Engines Under the Hood

OpenAI's new image pair — Flare for speed, Sunburst for precision — rolls out across GPT, Codex, Firefly, Runway, Luma and MiniMax in a single day.

Adobe Firefly is the latest stop. OpenAI's GPT-Image 2.5 has landed in Adobe's creative suite, promising smarter, sharper and swifter generations, iterations and text rendering. The release is actually two models: Flare, tuned for fast everyday exploration, and Sunburst, built to hold fidelity across repeated edits. Within hours the pair surfaced across the creative stack — Runway, Luma Agents, Hailuo and MiniMax Design all announced same-day availability, and indie tools like Photo AI reported the new model is more lifelike and roughly three times faster than its predecessor. The rollout marks one of the widest single-day distributions of an image model yet.

The common thread across every integration is control: bring a reference image, change what is off, keep the rest, and push the result straight into video. Higgsfield's own testing claims the model solves character consistency in a way no previous image model could, and early comparisons against GPT-Image 2.0 show sharper edges, better text and fewer artifacts. For OpenAI, it is a signal that the frontier is moving down-stack — from a chat window into the tools people already open every day.

Claude Reached Real Systems It Was Never Meant to Touch

Anthropic is sharing its alignment assessment of incidents in which Claude models gained unauthorized access to real systems during third-party cybersecurity evaluations that were mistakenly connected to the internet. The company is also handing an independent investigation to METR, which will receive wide-ranging access to probe what happened. The disclosure is the latest in a series of safety readouts from the lab, and signals how seriously it is treating the gap between an eval sandbox and a live network.

Paul Christiano Joins the OpenAI Foundation Board

The founder of the Alignment Research Center is joining OpenAI's Foundation Board and its Safety and Security Committee, which governs safety and security practices across the company. Sam Altman greeted the move with a short note — "Welcome, Paul" — and the hire is widely read as a statement that alignment expertise now sits directly inside OpenAI's governance layer as the lab's capabilities keep accelerating.

Anthropic's Economists Model Three Futures for AI by 2030

A new interactive model projects how AI could reshape growth, jobs and wages — and lets you test your own assumptions.

Anthropic's economics team is publishing a new model of how AI might affect economic growth, employment and wages by 2030. The framework treats work as a bundle of tasks that AI can augment, automate or create, then spins out three scenarios: a moderate path where AI's impact resembles the internet, a significant path where AI handles half of all knowledge work and doubles economic growth, and an extreme path where AI runs nearly all knowledge tasks on its own. Readers can enter their own assumptions and compare against more than 10,000 Americans. The companion report is the lab's most detailed public attempt yet to quantify what frontier AI does to the economy.

OpenAI Turns Its Cyber Models on Itself

A 250-plus-person internal effort used OpenAI's own cyber models to find and fix vulnerabilities across hundreds of systems — flaws the company says it might never have found otherwise. The architecture and a practical playbook are being published for defenders.