YouTuber PewDiePie says OpenAI suspended his account twice while he was developing Ajax, a small artificial-intelligence model intended to run on a user’s own computer. The project is designed to work inside Odysseus, a free, self-hosted application he launched in June, and is meant to handle tasks such as browsing the web and sorting email.
According to Decrypt, PewDiePie described the bans in a new video on his YouTube channel. Tech City Authority found no public response from OpenAI regarding his account or the allegations.
Ajax is reportedly a nine-billion-parameter model based on Alibaba’s open-source Qwen 3.5. It has not yet been made available for download. A countdown on the project’s download page had pointed to 08:25 Japan time on October 3, although the timer was later removed.
Why PewDiePie says he was banned
PewDiePie said he initially wanted to use OpenAI’s GPT-5.6 Sol to help create training material for Ajax. OpenAI released Sol, described in the source material as its flagship model, on July 9.
His stated goal was to distill knowledge from the larger system into a smaller model. In AI development, distillation involves training a less powerful model using answers generated by a larger one. PewDiePie said he was particularly interested in Sol’s reasoning tokens—the hidden text produced while the system works through a problem. OpenAI keeps those reasoning records encrypted.
He said he located a study describing how the encrypted material could be accessed through OpenAI’s own application programming interface. PewDiePie said he did not break OpenAI’s systems, but that his account was nevertheless banned. He disputed the decision and said the account was later restored.
After the restoration, he said he used Sol again to produce seed data, or initial examples for training Ajax. The account was then banned a second time. OpenAI’s terms of use prohibit using its output to develop competing models, a restriction that provides the stated context for the second suspension.
The issue comes amid wider attention on attempts to recover concealed reasoning data. Researchers reported in August that encrypted reasoning blocks from OpenAI, Anthropic and Google could be replayed to weaker related models, which then reproduced the material in plain text. They decoded 315,320 blocks from public code repositories. On September 30, OpenAI said it had disrupted a campaign linked to people associated with Moonshot AI, the company behind Kimi, to extract hidden reasoning from its models.
Ajax is designed to run locally
Unlike a cloud-based assistant, a local model runs on the user’s own hardware. PewDiePie’s project is intended to keep information such as emails and calendar data on that computer rather than sending it to a company’s servers. It also avoids a subscription charge, although the user’s machine must provide the computing power.
The reported nine-billion-parameter size places Ajax among the smaller models being developed for local use. Parameters are adjustable values used by an AI model to process information; in general, larger models have more capacity, although size alone does not determine performance.
PewDiePie presented Ajax as the first version of the project and said further work remains. He still needs to repeat the process used to remove refusals, compress the model through quantization and benchmark it to determine whether the changes damaged its performance.
He compared the resource demands of large and small models using his own rough calculations. He said one copy of a rumored trillion-parameter frontier model would require about 27 of his computers and electricity equivalent to roughly 150 homes. The comparison was offered as part of his argument that compact models can be more practical and efficient for personal computing.
Removing built-in refusals
PewDiePie used Heretic, an open-source tool, to make Ajax less likely to reject requests that assistants such as ChatGPT would refuse. The technique, known as abliteration, compares a model’s responses to harmful and harmless prompts, identifies patterns associated with refusals and removes them.
He acknowledged that the process had drawbacks, saying Ajax suffered “a little brain damage.” He estimated that the model still completed tasks such as inbox organization and web browsing successfully about nine times out of 10.
PewDiePie said he selected which refusals to remove and set a boundary around causing harm to other people or to oneself. He also said his lawyer advised him to make clear that Ajax was not designed to provide dangerous, actionable instructions.
The model is still being trained with GRPO, a method in which the system attempts a task 16 times and learns from the attempts that succeed. Whether Ajax will meet its stated goals will depend on the results of the remaining training, compression and testing, as well as its eventual release.
Moneyiar market picture
Moneyiar data at 09:43 on October 5, 2026, showed Ethena (ENA) at $0.2535012654, up 6.67% over 24 hours but down 1.78% over seven days. Ethereum Name Service (ENS) was $6.9126311671, rising 1.57% in 24 hours and 4.81% over seven days.
Compound (COMP) traded at $24.8537253173, down 0.66% over 24 hours and up 7.52% over seven days. Rain (RAIN) stood at $0.0121571707, falling 2.4% over 24 hours and 3.25% over the past week.
