Karpathy has tested whether Opus 5 can replace game studios

Signals Inbox·August 3, 2026·AI Creative Tools

Andrej Karpathy gave Opus 5 one paragraph from The Lord of the Rings, a one-million-token budget worth roughly $10, and one job: turn it into a 3D world. Two hours and about 5,500 lines of code later, he had a rough interactive scene. It is nowhere near a finished game, but $10 is suddenly enough to test work that once needed several specialists.

Q1What actually happened?

In his original post, Andrej Karpathy says he gave Opus 5 the opening paragraph of The Lord of the Rings and asked it to recreate the passage as a procedural Three.js world. He allowed the model to spend up to one million tokens, worth about $10. Opus worked for roughly two hours and produced around 5,500 lines of code for an interactive 3D scene.

Q2Did it really build a game?

It built something closer to a playable prototype than a proper game. The result had a 3D landscape, moving characters, buildings, effects, and basic interaction, but it was also visibly rough. There was no polished combat system, deep story, professional art, or months of testing. The important part is that one model connected many different pieces into a working world from almost no instructions.

Q3Why does the $10 matter?

Because early game production normally uses several expensive skills: programming, environment design, animation, technical art, and testing. Opus did not match professional workers, but it produced a usable experiment for less than the price of lunch. That changes how many ideas a small team can test. Instead of carefully choosing one prototype, a developer could generate ten rough worlds and keep the strongest one.

Q4Can Opus replace game studios now?

No. Finished games require consistent art, strong controls, good performance, fun mechanics, original direction, and thousands of careful decisions. Opus still struggled to inspect its own 3D output. It had to take screenshots slowly and could miss obvious visual problems. It can already compress parts of prototyping, but replacing an entire studio is a much harder test than producing one impressive browser demo.

Q5What is different from earlier AI game demos?

Many earlier systems generated short video-like game frames, simple levels, individual assets, or NPC dialogue. Karpathy tested whether one general model could write the actual software connecting a whole experience. It created ordinary Three.js code rather than dreaming every frame with a world model. That makes the result less magical, but more editable. A human developer can inspect the files, change the rules, and continue building from them.

Q6What is the bigger shift?

Games may slowly move from fixed products made for millions of people toward temporary worlds made for one person. Karpathy describes a future closer to an on-demand, personalized GTA. We are not there yet, but the cost of testing that idea has collapsed. The key question is no longer whether AI can generate one asset. It is whether a model can keep working long enough to combine code, art, movement, and rules into one coherent experience.

Q7So should game studios worry?

They should watch the prototype stage first. The near-term threat is not an AI instantly making the next Grand Theft Auto. It is a tiny team exploring ideas at a speed that once required a much larger group. If models become better at seeing their mistakes, testing gameplay, and keeping large projects consistent, the same cost collapse could move from demos into real production. This experiment shows the direction, not the finished destination.