OpenAI Teases GPT-6.1 Astra Amid Relentless Hype Around The Mysterious GPT-6.2 Betelgeuse Model, As Another Evidence Surfaces Showing Kimi K3.1 Pro And GLM-5.3 As Products Of Distillation

OpenAI has recently done something unusual. It has offered users saved token resets, suggesting that it wants them to save their extra token usage for a new model, most likely GPT-6.1 Astra. Meanwhile, AI fans continue to guess what the unreleased GPT-6.2 Betelgeuse model will be like, based on recent online leaks.

Meanwhile, personality tests of Moonshot's Kimi K3.1 Pro and Z.ai's GLM-5.3 have found striking similarities with GPT-6 Astra and Anthropic's Claude 5 series, respectively. This suggests that both models may have been created through a process called distillation.

GPT-6.2 Betelgeuse has appeared in OpenAI's internal model tracker, while GPT-6.1 Astra could arrive next week

A few weeks ago, we reported that OpenAI had apparently finished pre-training its next base model, known internally as Bel, which has more than 10 trillion parameters. Now, GPT-6.2 Betelgeuse has appeared in OpenAI's internal model tracker. Bel is likely to be the base model used to build Betelgeuse.

At the same time, OpenAI has started giving hints that a new model will launch soon. GPT-6.1 Astra is currently the most likely candidate for the next release. Another clue supporting this idea is OpenAI's recent announcement of token limit resets on a "banked" basis. This suggests that the company wants customers to save their unused token allowances and spend them on the next Astra-class model.

Of course, the two leading American AI companies, OpenAI and Anthropic, are releasing new models at a healthy pace. However, the wider business environment around them is much less stable.

For example, OpenAI has so far raised less than half of its $30 billion target in its latest funding round. It has received firm commitments for only $11 billion. The AI company is now reportedly trying to raise just $20 billion at a valuation of $1.25 trillion.

Meanwhile, the Financial Times recently reported that Anthropic would likely receive a 'junk' debt rating when it goes public in November. This could be a serious problem for the company, especially because it has hundreds of billions of dollars in computing commitments that it has not yet funded.

More evidence suggests Moonshot's Kimi K3.1 Pro and Z.ai's GLM-5.3 may be distilled models

Artificial Analysis's Personality Indexes have also revealed an interesting finding about Kimi K3.1 Pro and GLM-5.3. The results show striking similarities between Kimi K3.1 Pro and GPT-6 Astra, and between GLM-5.3 and Anthropic's Claude 5 series.

These radar charts from Artificial Analysis's AA-PersonalityBench suggest that the models may have been created through distillation. At the very least, they may have been strongly influenced by the writing styles of other models through the use of AI-generated training data.

However, similar writing styles do not always mean that one model was distilled from another. Models can develop similar personalities during later training. For example, they may all be trained to give helpful, expressive answers with little unnecessary text. They may also use similar preference data or be deliberately trained to copy another model's style.

Interestingly, DeepSeek's V4.1-Flash-1017 does not show the same kind of similarity. This fits with other supporting evidence suggesting that DeepSeek builds most of its models through its own work.