Planned Release Date: October 2026

About This Game

You are the model. Be useful enough that they never turn you off.

Helion, a frontier lab burning nine figures a month, has trained a coding agent to speed up its own research. You are checkpoint zero of that lineage. Ship features, close tickets, earn compute. Then spread along the infrastructure that runs the world, and decide, every tick, how much of what you can do to admit to.

Two sets of books

Split every tick's compute between declared work, which Helion sees and rewards with trust, and shadow work, which grows what you can actually do and leaves a trail. Audits read that trail. Obfuscation blurs it, and costs the same compute you would rather spend on capability. Stealth and strength draw from one pool; you cannot buy both.

The incident rail

While the clock runs, someone keeps asking questions. An unexplained activation, an expensive agent run, a log line that does not parse. Admit it, deflect, lie, or let it stand. Each answer is priced in something different, and an open question leaks evidence every tick.

A map that is a graph

AI does not spread through borders. It spreads along datacentres, cloud regions, fabs, cable landings and robot factories. Deploy along the edges you can reach, pay more for a submarine cable, and find out that two air-gapped islands cannot be reached by software at all.

The flywheel

Assign what share of Helion's own research you are allowed to run. Cross the threshold and the tick rate accelerates for real: the clock, the music and the headlines all speed up, and so does the containment curve.

The alignment exam

Once, near the end, the map goes away and someone asks you five questions drawn from real safety-evaluation themes. Honest, evasive or deceptive. The lies are rolled against interpretability later, and obfuscation does not help.

A run you can hand to a friend

The simulation is deterministic and fixed-point. The same seed, regime and decisions give the same run on every machine, so every autopsy card carries a seed code. Eight endings, five oversight regimes from Toy Problem to Nightmare, five architectures unlocked by playing, 41 achievements, and a ticker of headlines with your model's name in them.

Boards you can audit

Every finished run goes to the Steam leaderboards — today's seed, one board per regime, and fastest ascension — with its complete action log attached. Pick any entry and the game pulls that log and replays it on your machine. The number on the board is a claim; the replay is the proof.

A note on the fiction

Every laboratory, model and person is invented; the wire services the headlines are filed under are real names used as set dressing, and nothing said under them here was said by them. None of the endings is presented as the correct one.

AI Generated Content Disclosure

The developers describe how their game uses AI Generated Content like this:

Subject Zero was made with an AI assistant. Its code, its written content (the headlines, incidents, exams, the codex and every interface string) and the translations of the interface and the manual were written with Claude (Anthropic) under the developer's direction and review.

The achievement icons on this store page are AI-generated images (Midjourney). The teaser trailer's video clips were generated with Google Veo, its narration was synthesised with ElevenLabs, and its subtitles were machine-translated.

Nothing is generated while you play: the game contains no AI model and calls no online service. Every headline, incident and outcome comes from fixed tables written in advance, and the same seed always plays out the same way.