Spud, OpenAI's upcoming model expected to ship as GPT 5.5 Pro, appears to have already been quietly pushed to some users under the hood of GPT 5.4 Pro. People started noticing something felt different. Faster responses, different output styles, noticeably better quality. An OpenAI employee confirmed in replies that the team was actively monitoring feedback, which is about as close to a confirmation as you tend to get before an official announcement.

The leaked outputs circulating online are hard to ignore. In one test, the model generated a fully playable Minecraft-style game in a single prompt, with working physics, inventory systems, and a world that actually renders properly. Previous models have attempted this kind of test and produced something technically functional but barely playable. This one is a different category of output. There is also an Xbox controller rendered entirely in code, photorealistic enough that people assumed it was an image until they saw the source. A full Pokemon battle game built from a single screenshot. A 3D animated pelican riding a bicycle through a voxel world, complete with a lighthouse, sun rays, and a crab, all generated in one shot.

None of this is official. No benchmarks, no release notes, no blog post. Just outputs that keep showing up and a growing consensus that what people are testing right now is not the same model that was there a week ago. Greg Brockman, one of OpenAI's co-founders, has called the upcoming release the culmination of two years of research and an early version of what is still to come.

If these leaked generations are real, Anthropic's timing with Opus 4.7 is about to get very complicated.

Chronicle is a new experimental feature inside Codex that takes periodic screenshots of your Mac throughout the day and uses them to build a running memory of what you have been working on. The idea is that you should never have to re-explain your context. You type "why is this failing?" and Codex already knows what you were doing, what broke, and where to look.

Sam Altman's reaction to the internal demo was telling. The working name for this feature was "telepathy." He said it feels like it.

The people at OpenAI who have been using it say it has changed how they work with Codex entirely. That is a strong endorsement from people who could use any tool they want. The caveat is that it is early, it consumes a significant amount of tokens, and the screenshots are stored as unencrypted files on your Mac, which OpenAI flags upfront as a privacy consideration worth knowing about before you opt in.

The broader picture is that Codex is quietly becoming something different from what it launched as. Memory, background computer use, 90-plus integrations, and now a feature that watches your screen so you never have to restate context. Whether that vision of an always-on work agent is exciting or unsettling probably depends on who you ask. Either way, it is where OpenAI is going.

Anthropic shipped a new feature in Cowork today called Live Artifacts. Claude can now build dashboards and trackers that stay connected to your apps and files, so every time you open one it refreshes with current data. Version history is saved automatically, meaning you can come back a month later and pick up exactly where you left off.

It is a small announcement compared to everything else this week, but the direction is clear. Claude is quietly becoming the place where your data lives, not just the tool you ask questions to.

Available now on all paid plans.