SoftBank’s Masayoshi Son Says AI Is Now Building OpenAI’s Next Model

SoftBank CEO Masayoshi Son told CNBC Friday that an existing OpenAI model is actively designing its successor. He called the development a clear signal that artificial superintelligence has nearly arrived.

Son Moves His ASI Timeline Forward Sharply

Son said he had spoken directly with OpenAI CEO Sam Altman and engineers at the firm. They confirmed that a current model is being used to architect the next generation of model. Once that cycle takes hold, Son argued, the trajectory becomes exponential. Engineers will simply no longer possess the capability to keep up. “That’s a superintelligence,” he told CNBC.

The SoftBank chief acknowledged he had previously set a 10-year window for ASI, defining it as AI roughly 10,000 times more capable than any human. He now says that estimate was deliberately cautious. His private view at the time was closer to four years. Today he puts the arrival date at roughly two years from now. He also told CNBC he personally uses ChatGPT for two to three hours each day, because he considers it already smarter than he is across most topics.

Background: OpenAI Has Already Flagged Self-Referential Development

OpenAI has not been silent on the subject. In February the company noted that its GPT-5.3-Codex model played a meaningful role in its own creation. The Codex team used early versions of the model to debug training runs, handle deployment tasks, and assess evaluation results. An OpenAI spokesperson declined Friday to comment on unreleased models but pointed to those existing examples of AI-assisted development.

A separate OpenAI research paper published this month described what it called “early signs” of recursive self-improvement in current systems. The paper flagged potential governance challenges for which institutions are not yet prepared.

Anthropic Urges Caution as the Stakes Rise

The optimism from Son landed the same week that Claude developer Anthropic issued a pointed warning. In a blog post, the company described recursive self-improvement as a process where an AI system independently designs its own successor. Anthropic acknowledged potential upsides but stressed that unchecked progression raises serious risks of humans losing meaningful oversight. The firm called for coordinated action among major AI laboratories to deliberately slow that development down.

Whether Son’s description of OpenAI’s model-building process technically constitutes recursive self-improvement remains unclear. What is clear is that the gap between cautious industry warnings and bullish investor forecasts is narrowing fast.

Read Next: OpenAI Launches GPT-5 as Its Most Capable and Unified Model Yet