News
What's happening in AI game development.
Announcement
Open-source tool trompe paints a gray game blockout into a textured level
The README's village example used 128 painted views and cost $4.44 in 26 minutes. Input must be a UV-mapped .glb model in Y-up orientation.
Press release
BlueprintsLab says Ultimate Engine CoPilot V2 gives Unreal agents 1,655 editor tools
The Unreal Engine 5.5–5.8 plugin claims agents that compile, playtest and review their own changes. Plans start at $25 a month; the feature claims come from the company's press release.
Announcement
Google DeepMind unveils Genie 4 with native audio and 2-minute consistent worlds
Users can change the weather, camera or objects with text while a world runs. For now it's a research preview for trusted testers, with a waitlist for Gemini API partners.
Announcement
anymove extracts game-ready creature animation from a single AI-generated video
The open-source tool handles wings, tails or 8 legs and exports a 24 fps GLB, but training needs a 24 GB NVIDIA GPU. A 5-second source clip costs about $0.26.
The best of the week, every Friday
The top stories and the AI Game of the Week in one short email.
Announcement
Epic opens AI MegaJam with $65k in prizes after saying AI is optional
Teams of up to 5 people, plus any number of agents, get 4 days to build an Unreal or UEFN game. Each track also offers continuation grants of up to $50,000.
Announcement
Hi3D releases v1.0 open weights to refine rough meshes into high-res geometry
The model, formerly Sparc3D, refines geometry up to a 1536³ grid but makes no textures, and its top setting has only been validated on a 96 GB GPU.
Unity brings its official plugin to Grok Build, joining Claude Code and Codex
Grok Build gets the same 30-plus Unity skills, from Shader Graph to multiplayer, for Unity 6 and above. Unity says usage data on its CLI and plugins is due by the end of October.
Sakurai warns AI will flood game stores, and Miyamoto questions what "creating" means
Sakurai expects store releases to keep climbing past 19,000 with most "unplayable." Miyamoto sees AI as a reason to rethink creating versus working.
Ludo.ai brings asset generation to the terminal and adds 32 free agent tools
The free Asset Toolbox needs no account or credits and handles engine prep such as pivots, colliders and loop checks. Generation itself stays on paid plans.
Digital Sun cancels public ReVamp playtest as agent-built Terraria rewrite goes live
ReVamp's first test moves to a small Discord group over cloning fears. The same day, a Terraria web rewrite built in 70 days with 123.8B Claude Code tokens appeared on GitHub.
Report: Decomp.dev shuts its project index, saying AI broke the decompilation community
Existing project pages and historical data stay online, but new submissions are closed. The site's creator says AI can speed up the work; the damage is to how volunteers coordinate.
1 in 3 Steam Next Fest demos disclose AI as players request filters
SteamDB snapshots put disclosure at 36.4% of more than 7,000 October demos, though one count of the disclosure banner alone found 14.7%.
Microsoft and Epic file DMCA notices against agent-built browser game recreations
The takedowns hit ports of Minecraft, Halo and Fortnite. Tim Sweeney says Epic is working to drop a EULA rule limiting AI training on Unreal and MetaHuman outputs.
Announcement
Reka says its 19B Rho-1 model can run real-time, steerable game worlds
Demos show photoreal, voxel and low-poly worlds from one checkpoint, but longer streams drift in layout. It's a research preview with no public API or open weights.
Announcement
6 open-source tools let coding agents playtest and check the games they build
Ludra, Glitchless, PlayCanvas Stage and new plugins for Unity, Godot and Unreal arrived between October 21 and 28. Most are new and have few GitHub stars so far.
Report: Ubisoft makes its Motherbrain generative AI team a company-wide unit
Led by Neo NPC R&D director Xavier Manzanares, the team will reportedly build shared generative-gameplay middleware for every studio. Ubisoft hasn't confirmed it.
Announcement
Viggle says its JST-2 world model builds 3D splat scenes in about a second
Scenes are static and physics is rule-based for now. The Scene API charges $0.06 to prompt a scene and $0.01 for other tasks.
Announcement
KanaWorks_AI's Audio Studio MCP lets coding agents generate music and SFX locally
The MIT-licensed tool runs Stable Audio 3 on Apple Silicon or NVIDIA GPUs, with no cloud subscription. Version 1.0 adds a browser workbench for comparing and editing takes.
Announcement
Genex releases open-source desktop app that runs AI agents as a game-dev team
The free, MIT-licensed app has Claude Code, Codex or local models plan, build and judge Three.js browser games. It runs on macOS and Linux, with a Windows preview build.
Announcement
Open-source Blender skill packs let agents build node graphs and export to engines
EL4CTEO's 22 skills export to Roblox, Unity, Unreal and Godot, while gokayfem's tools rebuild videos and Gaussian splats as editable node graphs.
Announcement
Coding agents ship native ports of Tears of the Kingdom, Portal 2 and GoldenEye
Most of the ports need the player's own game files, and their legal status is unsettled. A co-op playbook is also adding multiplayer to Deus Ex and other single-player games.
Developer builds 61 KB WebGPU ocean with an agent that screenshots its own work
The creator says Claude Code and Opus 5.5 built it in about 30 minutes. Clones of the Moon, Mars and a neon city followed, one adding a separate critic agent.
Report: Agents wrote Pixkrieg's 3,700 commits under strict CI and a review agent
An October 18 Codex Insider article says agents built the browser RTS in 2 weeks under strict linting, 90% test coverage, mutation testing and a review agent. AIGamer couldn't confirm it with the developer.
Research
AnyGameRenderer model repaints gray-box gameplay footage into polished visuals
The research model keeps each entity's position and timing, but it still misses about 1 in 5 changes in entity counts and works on recorded video, not live games.
Research
Roblox's OpenGameEval tests 13 frontier models in Studio, and the best solves 51.7%
Agents that inspected the scene before editing passed up to 13.4 points more often, and 6 of the 84 tasks went unsolved by every model.
