GTA VII is an open-world game being written entirely by Claude, starting on 19 November 2026 and targeting release on 19 November 2046. It begins from an empty repository with $200 in API credits. There is no studio and no human-written game code: Claude plans the work, writes the code, builds it, tests it, fixes what breaks and commits the result.
The premise is simple. GTA VI arrived roughly thirteen years after GTA V. Each generation of open-world game is bigger than the last, so the next one plausibly takes twenty. If the clock is twenty years, the only sensible time to start is now.
- Build page: the live terminal, the process strip showing which step Claude is on, and the dashboard of progress, budget and module status.
- Roadmap: the seven-phase plan from 2026 to 2046, written out in full.
- Docs: this page.
- Small steps. Every change is small enough to build and test in minutes.
- Tests before trust. Nothing is merged without a passing test.
- Never rewrite what works. Twenty years of budget can't survive repeated rewrites.
- Everything in the open. Thinking, failures and fixes are all shown.
You don't need to install anything to follow the build.
- Open the build page and scroll to the terminal.
- Watch the process strip above it. The highlighted step is what Claude is doing right now.
- Click into the prompt at the bottom of the terminal and type
help. - Try
status,budgetandmodulesto see where things stand. - Use pause to hold the feed while you read (it catches up when you resume), and 1x / 2x / 4x to change how fast commands are typed out.
- Open the code page to read every file Claude has committed.
- Switch the site between light and dark with the button in the top bar, and the terminal on its own with the light / dark button in the terminal bar.
The whole team is one model: Claude Fable 5.2, Fable's latest model. It does every job a studio would normally split across departments.
| Role | Done by | What that means here |
|---|---|---|
| Engine programming | Claude Fable 5.2 | Job system, memory, rendering, physics, streaming. |
| Gameplay programming | Claude Fable 5.2 | Vehicles, NPCs, police, combat, economy. |
| Tools | Claude Fable 5.2 | Command-line tools in place of an editor. |
| QA | Claude Fable 5.2 | Writing and running tests, soak tests and automated playthroughs. |
| Writing | Claude Fable 5.2 | Story, missions, dialogue, radio. |
| Production | Claude Fable 5.2 | Quarterly reviews, roadmap updates, budget tracking. |
When a newer model is released, the team line is updated and the change is recorded in the decision log.
All work happens in one repeating loop. Each pass handles one small task, such as "add tire friction" or "fix the pedestrian pathing bug". The process strip on the build page lights up the current step.
| Step | What happens |
|---|---|
| Pick | The next task is taken from the current roadmap phase, smallest useful task first. |
| Think | Claude reasons about the task: what exists already, what the simplest correct version is, what could break. This is the italic thinking line in the terminal. |
| Write | Code and tests are written. Tests for new behaviour usually come first. |
| Build | The target is compiled. A failed build goes straight back to Think. |
| Test | The tests for that target run: basic, load, edge and stress cases. |
| Fix | Only when a test fails. Claude reads the failure, reasons about the cause, patches it and re-runs the tests. |
| Commit | Passing work is committed with a message naming the module and the change. |
GTA VII is a single C++ monorepo built with CMake. It has its own engine rather than a licensed one, so every system can be written and tested by Claude end to end.
gta7/ ├── CMakeLists.txt ├── README.md ├── docs/ │ ├── decisions/ # one file per architectural decision │ ├── reports/ # quarterly progress reports │ └── parked/ # tasks set aside after repeated failures ├── src/ │ ├── engine/core/ # jobs, memory, entities, time │ ├── render/ # renderer and backends │ ├── physics/ │ ├── world/stream/ # map, terrain, streaming │ ├── ai/npc/ │ ├── vehicles/ │ ├── audio/radio/ │ ├── missions/ │ └── online/ ├── tests/ # mirrors src/, one suite per module ├── assets/ # source assets and pack rules └── tools/ # command-line tools in place of an editor
Lower layers never depend on higher ones. This keeps tests fast and lets each layer be finished and frozen in turn.
missions ──► ai/npc, vehicles, audio/radio ai/npc ──► world/stream, physics vehicles ──► physics, render world ──► render, engine/core physics ──► engine/core render ──► engine/core online ──► everything above, read-only
These are the modules shown in the modules card on the build page. Each one has its own tests and its own progress bar.
| Module | Responsible for | Phase |
|---|---|---|
repo/scaffold | Repository layout, README, licences, coding standards. | 0 |
build/cmake | Build presets, CI pipeline, sanitizers, packaging. | 0 |
engine/core | Job system, fibers, entity component system, memory arenas, time, save games. | 1 |
render | Graphics backends, lighting, materials, shadows, sky, weather visuals. | 1 – 2 |
physics | Collision, rigid bodies, ragdolls, buoyancy, destruction. | 1 – 3 |
world/stream | Terrain, roads, districts, interiors, level of detail, streaming. | 2 |
ai/npc | Pedestrians, traffic, police, schedules, reactions to the player. | 3 |
vehicles | Handling, damage, sound, cars, bikes, boats and aircraft. | 3 |
audio/radio | Mixing, music stems, radio stations, DJs and ads. | 3 – 4 |
missions | Scripting, checkpoints, cutscenes, dialogue, heists. | 4 |
online | Servers, sessions, persistence, co-op, anti-cheat. | 5 |
Every commit is built and tested automatically. A commit that breaks any test is reverted.
| Test type | What it checks | When it runs |
|---|---|---|
| Basic | The feature does what it should in the normal case. | Every commit |
| Load | It still works with many instances at once. | Every commit |
| Edge | Boundary cases: zero, maximum, empty, overlapping. | Every commit |
| Stress | Long runs and random input to shake out crashes. | Every commit, short; nightly, long |
| Performance | Frame time stays within its budget (16.6 ms for 60 fps). | Every commit |
| Determinism | Recorded replays play back identically, frame for frame. | Nightly |
| Soak | Claude drives and plays through the world for hours. | Nightly, from Phase 3 |
| Playthrough | Every mission is completed automatically. | Weekly, from Phase 4 |
In the terminal, test results appear as [PASS] or [FAIL] followed by the test name and either its run time or the reason it failed.
- Found. A test fails. The wanted level goes up one star.
- Diagnosed. Claude reads the failure and writes its reasoning as a thinking line.
- Patched. A patch line shows the file, line and size of the change.
- Verified. The same tests re-run and pass. The wanted level drops one star.
- Committed. The fix is committed with a message starting
fix. - Regression test. The failing case is kept in the suite so it can't come back unnoticed.
The project starts with $200 in API credits. Every call Claude makes costs tokens, and the budget card on the build page tracks real spend against that $200, worked out from the token counts the API returns.
- Each task logs its token count and cost in the terminal, on the dim line after the commit.
- The api budget card shows total spent, percentage of the $200 used and total tokens.
- Type
budgetin the terminal for spent, remaining and a bar.
The starting credits cover Phase 0 and the first months of Phase 1. The breakdown is on the roadmap page.
Later phases are funded by top-ups. Spend slows as the budget runs down, so the remaining credits always go to the highest-priority work.
With a small budget and a long timeline, the most expensive mistake is spending money going in circles. These rules prevent that.
| Guardrail | Rule |
|---|---|
| Task cap | Each task has a maximum dollar cost. Reaching it parks the task. |
| Retry limit | Two failed fix attempts park the task and revert its changes. |
| Hard stop | When the total budget is spent, the loop stops until it is topped up. |
| Sandbox | Generated code is compiled and run with time and memory limits, in an empty folder. |
Each kind of line has its own colour so you can scan the feed quickly.
| Line | Example | Meaning |
|---|---|---|
| Thinking | thinking Tests first... | Claude's reasoning before it acts. |
| Command | $ ctest -R ... | A command Claude runs. |
| Write | write src/... | Files being written and how many lines. |
| Pass | [PASS] ... | A test passed. |
| Fail | [FAIL] ... | A test failed, with the reason. |
| Patch | patch src/... | A fix: file, line, lines added and removed. |
| Commit | [main 3f9a1c2] ... | Work committed to the main branch. |
| Cost | tokens +12.4k ... | Tokens used by the task and total spend. |
| Button | What it does |
|---|---|
| light / dark | Switches only the terminal between a light and a dark theme. |
| pause / resume | Stops the feed so you can read, then continues it. |
| 1x / 2x / 4x | Changes how fast commands are typed out. |
| clear | Clears the screen. The build keeps going. |
The feed scrolls automatically. If you scroll up to read, it stays where you are until you scroll back to the bottom.
Type these into the prompt at the bottom of the terminal and press Enter.
| Command | Output |
|---|---|
help | Lists every command. |
status | Current day, progress, current task and counters. |
budget | Spent, remaining, tokens and a budget bar. |
eta | Release date and time remaining. |
modules | Progress bar for every engine module. |
roadmap | The seven phases and which one is running. |
team | Who is building the game. |
wanted | Current wanted level and open bugs. |
ls | Top-level files in the repo. |
log | The latest real commits. |
files | Every source file and its line count. |
cat <path> | Prints a file from the repo. |
theme light / theme dark | Switches the terminal theme. |
pause / resume | Stops or restarts the feed. |
speed 1 / 2 / 4 | Sets the feed speed. |
docs | Opens this documentation. |
about | A short description of the project. |
clear | Clears the screen. |
Every number on the build page, and how it's worked out. Counters are real. They come straight from the build server: every token is counted from the API response, every line from the git repo, every pass and fail from an actual compiled test run.
| Metric | Definition |
|---|---|
| Progress to ship | Time elapsed since the first commit as a share of the 20 years to release. |
| Elapsed | Days, hours, minutes and seconds since the first commit. |
| Time to release | Countdown to 19 November 2046. |
| API budget | Credits spent out of the starting $200, with total tokens. |
| Lines written | Lines of C++ and tests in the repo right now. |
| Tests passed | Total passing test cases across every real test run. |
| Bugs fixed | Failures that were patched and verified. |
| Wanted level | Failures that are still being worked on, zero to five stars. |
| Modules | Features committed against the planned backlog for that module. |
A nod to the game itself. Every failing test adds a star; every verified fix removes one. Five stars means five bugs are loose at once.
| Stars | Meaning |
|---|---|
| 0 | Clean. Everything passes. |
| 1 – 2 | Normal. Claude is fixing something. |
| 3 – 4 | Busy. Several failures at once, usually after a big change. |
| 5 | Maximum heat. New features pause until the count comes down. |
There are two separate switches.
- Site theme: the button in the top bar switches the whole site between light and dark.
- Terminal theme: the light / dark button in the terminal bar, or
theme lightandtheme dark, switches only the terminal. You can have a dark terminal on a light page, or the other way round.
Both choices are remembered in your browser for next time.
Is this an official GTA game?
No. GTA VII is an independent fan project. It has no connection to Rockstar Games or Take-Two Interactive.
Why twenty years?
GTA VI arrived roughly thirteen years after GTA V, and each generation is bigger than the last. Twenty years is the planning estimate for a sequel at that scale.
Why only $200?
To show how far a small, disciplined budget can go when work is done in small tested steps and nothing is rewritten without a reason. Later phases are funded by top-ups.
What happens when the $200 runs out?
Spend slows as the credits run down, so the most important work is done first. Top-ups fund the following phases.
Which model is building it?
Claude Fable 5.2, Fable's latest model. It's the whole team.
Does a human write any code?
No. Claude writes all the game code, tests and tools. People set the direction and fund the work.
Will it be playable before 2046?
Parts of it. The roadmap has playable milestones along the way: a drivable car in 2032, the first full mission in 2038 and a public online load test in 2045.
Can I follow along?
Yes. The build page shows the work as it happens, and the roadmap is updated after every quarterly review.
| Term | Meaning |
|---|---|
| CI | Continuous integration: building and testing every change automatically. |
| Determinism | The same inputs always give the same result, so replays match exactly. |
| ECS | Entity component system: game objects are ids with data attached, processed in bulk. |
| Fiber | A lightweight task that can pause and resume, used by the job system. |
| LOD | Level of detail: simpler versions of objects used far from the camera. |
| Parked task | A task set aside after repeated failures, to be retried later. |
| Soak test | A long run, often hours, to find problems that only show up over time. |
| Streaming | Loading parts of the world as you approach them, with no loading screens. |
| Token | The unit API usage is measured and billed in. |
GTA VII is a fan project. Grand Theft Auto and related names are trademarks of their owners and are used here only to describe the project. This project is not affiliated with, endorsed by or connected to Rockstar Games, Take-Two Interactive or Anthropic. Claude is used as a tool to write the code.