Gamer set GPT-6 Astra to play through Fallout 3 – the AI reached the ending in 59 hours and started paranoidally checking saves

An enthusiast going by imjustnewatai ran an experiment putting the GPT-6 Astra AI model through the main story of Fallout 3.

Running on maximum settings through the Codex agent environment, the playthrough took roughly 59 hours of real time. The final edited video runs 4 hours and 32 minutes and went up on YouTube split into chapters.

The experimenter played through the opening himself. He then handed Codex a save right after escaping Vault 101 and gave it a simple task: get outside.

From there the model picked its own route and actions, working through the entire main story:

• heading to the Galaxy News Radio station

• searching for the main character father

• retrieving the G.E.C.K. (Garden of Eden Creation Kit)

• escaping Raven Rock

• launching Project Purity

Gear changed throughout the run. Against the behemoth near the radio station, the AI broke out the Fat Man, and it relied on hunting and assault rifles inside the vaults. By the final assault it had companion Fawkes, a plasma rifle, and power armor training.

By the last save, the character had only hit level 10. The Energy Weapons skill sat at 54, and the model picked Scrounger as a perk. That build kept ammo flowing in reliably, something the model needed since it kept running low.

Vault 87 turned out to be the toughest section. The AI died there several times, reloaded, ran short on ammo, and had to backtrack for healing. One recovery hinged on something small – the game needed to "run for a bit" before a bed would activate.

The model status update at that moment read:

Кровать заработала после того, как игре дали немного поработать. Здоровье полностью восстановлено до 260/260.

After that, the AI reports started sounding "paranoid," as the experimenter put it. The model kept double-checking whether it might lose progress. On a staircase in the Citadel, it wrote:

Я сверяю каждое короткое перемещение с видимой геометрией, чтобы уйти, не потеряв сохранённый прогресс.

After Raven Rock, another message showed up:

Я делаю резервную копию этой контрольной точки и проверяю запись перед встречей с Братством.

The enthusiast noted he uses "paranoid" jokingly. All the quoted messages are real though, and constant save-checking became a defining trait of the whole run.

In the end, the water purifier came online and the game rolled its ending. When the hero woke up in the Citadel, Elder Lyons confirmed clean water was flowing, and the "Take it Back!" quest got marked complete. The model never touched the FEV virus, and the Broken Steel story content wasn't part of the run.

Once it wrapped up, Codex summed things up with a core lesson – check the result first, then repeat the input only if needed. A shot blocked by geometry meant repositioning was needed. The bed needed a short pause to work. Pressing a keypad required confirming the input actually registered.

One of the model public updates sums up that approach well:

Подойдя ближе, я убрал препятствие, которое мешало с предыдущими выстрелами.

The model controlled the character through standard input and worked in cycles of pause, observe, act. It leaned on saves and loads, companions, and fast travel. Difficulty shifted throughout the run and sometimes dropped down to Very Easy.

The edited video cuts out the pauses where the model was reasoning, and sped-up segments are marked separately. The full run took 59.3 hours, with 8 hours of the original footage preserved before speed-up was applied. The previously mentioned figure of 56 hours was just a rough estimate, according to the experimenter.

Viewers noticed the AI leaned heavily on V.A.T.S. in combat. Real-time aiming gives the model trouble because of lag – picking a single dialogue line sometimes took around two minutes, and many segments in the video run at 4x speed. One viewer spotted a moment where the model actually fired manually, landing two shots in a row to make up for slow image processing.

Some commenters saw practical value in the experiment, pointing to game QA testing as an example.

Others floated the idea of AI opponents for turn-based strategy games like Civilization. A few users worried this kind of tech could lead to bots in MMORPGs that are indistinguishable from real players.