Capcom has put two pieces of its in-house engine out in the open so that language models have something to read.
That is the detail from 2 October I keep turning over. At the CAPCOM Open Conference RE:2026, in Tokyo’s Ariake Convention Hall, programmer Satoshi Ishida gave a talk with a title only an engine team would write — “The Outlook and Future of the REX Project, Further Evolving the RE Engine for the Next Generation” — and laid out a staged plan to grow the RE Engine into what Capcom calls an AI-generation game engine. Two of its subsystems, RE:Dox and RE:Log, have already been released publicly, and according to GameBiz’s report one of the reasons is AI training.
Sit with that for a second. The reason an assistant is useless inside a proprietary engine is that it has never seen one. Unreal has a decade of public documentation, forum answers, plugin source and tutorials behind it, so a model can say something sensible about a Blueprint node. An engine that exists only on a studio intranet has none of that, and no amount of general competence covers the gap. Capcom appears to have worked out that if it wants tooling a model can drive, it has to publish enough of the toolchain for the model to learn the shape of it.
REX has been coming since 2023
REX stands for RE neXt Engine, and Capcom first put the name in public in October 2023. The RE Engine is not being thrown away. REX is a staged modernisation of what is already there, which is the correct way to do this and the slow one — the alternative is a from-scratch rewrite that eats four years and ships nothing, and plenty of studios have tried it.
Ishida named three pressures behind the work. Games have grown large enough that iteration has slowed badly. The developer base has become more varied and needs tools that are easier for everyone to use. And AI has moved fast enough that getting it into the workflow has stopped being optional.
The first of those is the one I believe hardest. Anybody who has worked on a project as it crossed from big to enormous knows the moment when the build stops being something you can poke at and starts being something you submit a request to.
Five boxes on one slide
The presentation listed the project’s main systems:
- RE:Dox — converts mismatched data formats into one common structure for faster processing
- RE:UI — the interface layer for development tools
- RE:Log — logging and communication, collecting logs into a central store developers can search
- RE:Flows — scripting for users
- RE:Runtime — the game runtime itself
Every one of those is plumbing. There is no asset generator on the list, no texture synthesiser, nothing that makes a picture. The named targets are data handling, tool optimisation, code assistance, runtime processing, logging, and above all QA through automated test plays and bug detection.
Which is roughly where I have argued the technology earns its keep — the same conclusion I came to about where AI actually fits in a 3D pipeline, and it is reassuring to see an engine team arrive at it with a budget attached.
Machine-legible and junior-legible are the same thing
Here is the part worth stealing for your own studio, whatever engine you are in.
Everything on that slide that makes the toolchain easier for a model also makes it easier for a person who joined last Tuesday. One data format instead of nine. A consistent interface layer instead of twelve tools that each invented their own. Logs in one searchable place instead of scattered text files and someone’s memory of which flag to pass. Scripting a non-programmer can read.
Those are the things every in-house pipeline is supposed to have and most do not, because the work is invisible and nobody gets promoted for it. “The AI needs it” turns out to be an argument that gets the work funded. I will take it.
The line on finished assets, and where it sits
Capcom drew its boundary in public at the shareholder meeting of 23 March 2026: the company will not implement AI-generated assets in its game content. No generated faces, no generated backdrops, no synthetic voice acting. The phrasing it used was that fans should know what they see on screen is real artist work. Internally, it said, AI would be used to improve development efficiency, with methods being tested across the graphics, sound and programming departments.
That boundary is holding so far, and it sits exactly where the CESA survey found Japanese studios already were — heavy adoption, concentrated on admin and process rather than on the frame the player sees. Capcom is not out in front of its industry here. It is writing down what the industry is doing and committing to the half of it that is defensible.
Whether a line between tooling and assets survives contact with a deadline is a different question, and one nobody can answer from a conference slide. Code assistance that writes a shader is tooling by Capcom’s definition. The shader ships. I would want to know where that gets filed.
QA is where it bites first
Automated test plays and bug detection are the near-term prize, and they are a genuinely good fit. A machine that will replay the same traversal four hundred times overnight and flag the forty where the character fell through the floor is doing work no human should be asked to do, and it does not get bored and stop noticing on run two hundred.
It also only catches the broken, and for an environment artist the expensive failures are rarely broken. A seam in a lightmap, a LOD that pops in the player’s eyeline, a decal fighting the surface it sits on, a kit piece rotated a degree out — all of those pass every automated test ever written and all of them get caught by somebody walking the level with their eyes. Detecting those is a far harder problem than detecting a crash, and no part of Ishida’s slide claims to.
The conference ran on in Tokyo through 3 October, with an Osaka leg on the 17th and 18th. Capcom publishing the talks afterwards is how most of what we know about the RE Engine got out in the first place, and on current form the RE:Dox repository will tell us more about REX than the keynote did.