Tokenizer matches OpenAI's
A model named "fe-pichu" appeared on Oct. 27 (the same day in Japan time) on "Arena," a site that compares AI models by pitting them against each other anonymously. @aitrackerbot, an X account that tracks the appearance of models, reported that fe-pichu uses the same "tokenizer" as OpenAI's models.
A tokenizer is the method an AI uses to split text before processing it, and it differs by developer. Even for an anonymous model, examining how it splits text can give a clue to its developer. @aitrackerbot said it could be a small OpenAI model, while leaving open the possibility that it is one of Microsoft's "MAI" models, which use the same tokenizer.
Seen as "GPT-6.1 Luna"
On the night of the same day, @notjazii, who posts information on new models, speculated that fe-pichu could be "GPT-6.1 Luna." The assessment was that its output quality fell short of "GPT-6.1 Terra," a low-priced model released the day before, and that it was a small model.
OpenAI's models line up as the top-end "Astra," "Sol" for everyday work, the cheaper "Terra" and the small "Luna." In the GPT-6.1 generation, Sol was released on Sept. 29 and Terra on Oct. 26, while Luna has not yet been released.
Higher-end "fe-raichu" also appears
On Oct. 28, @aitrackerbot reported that "fe-raichu," which also uses OpenAI's tokenizer, had appeared. Based on the naming, it could be a higher-end model than fe-pichu, according to the account.
Both are guesses based on the tokenizer and impressions of the output, and no material indicating model names or timing has been found. OpenAI has not made any announcement.
"ika" also appears in November
On Nov. 3 (the same day in Japan time), @aitrackerbot reported that "ika," another anonymous model using OpenAI's tokenizer, had also appeared on Arena. @aitrackerbot analyzed that ika's behavior did not match any of GPT-5.5, GPT-6 or the open model gpt-oss, and that it was likely a new, unannounced OpenAI model. It did not indicate any relationship to fe-pichu or fe-raichu.
