OpenAI reportedly acquired Weights.gg, the startup behind the Replay voice-cloning app, in a quiet January deal that went undisclosed for four months, according to reports by the New York Times and The Information. The deal, which only came to light on May 15, deepens the lab’s voice-technology push while underscoring its cautious posture on synthetic speech. Neither company has confirmed the transaction, and financial terms remain unknown.
The Information reported as early as January that OpenAI had bought Weights.gg, acquiring its intellectual property and roughly a half-dozen employees. The outlet noted that OpenAI did not plan to integrate the startup’s consumer-facing products. On May 15, the New York Times provided the first public disclosure of the acquisition, citing people familiar with the matter. Neither report disclosed a purchase price.
Weights.gg operated a social AI creation platform—a single hub for generating images, music, chat and audio—but its Replay app, which let users clone voices with a short audio clip, drew the most notice. The company’s blog advertised free AI voice cloning in April 2025. The startup raised roughly $4 million in a seed round led by Kleiner Perkins, with Original Capital and Precursor Ventures also participating, according to PitchBook data cited by the Times. By late March 2026, Weights was winding down: its main services shut down on March 31, and a goodbye notice appeared the following day. Replay was explicitly excluded and may have continued operating.
The acquisition fits a multi-year voice buildout at OpenAI. In March 2024, the lab previewed Voice Engine, demonstrating it could generate convincing speech from a 15-second audio sample, but declined to release it widely, citing impersonation and disinformation risks. On May 7—about a week before the Times report—OpenAI launched GPT-Realtime-2, GPT-Realtime-Translate and GPT-Realtime-Whisper, expanding its realtime voice models for developers. Its usage policies, updated in October 2025, explicitly restrict voice likeness, consent and impersonation.
If confirmed, the deal is a small talent-and-IP grab, not a strategic pivot. It suggests OpenAI is acquiring voice talent at a moment when synthetic speech remains one of the most contested areas in AI safety. The timing—a January acquisition, a March shutdown, a May model launch—shows the Weights team was likely absorbed months before the models were announced, though the lab has not linked the two events and no evidence ties Weights’ technology to those releases. The transaction highlights a dual-track challenge: pushing voice capabilities forward while maintaining strict guardrails. Absorbing a team that built a consumer voice-cloning app could serve either purpose.
Several important details remain unconfirmed. The purchase price and exact deal structure are unknown. It is not clear which teams the roughly half-dozen employees joined, or whether any of Weights’ intellectual property will appear in future OpenAI products. The Information indicated that the product would not be integrated. OpenAI has not commented on any aspect of the reported deal. The handling of user-generated voice models and consent records from the Weights platform has not been disclosed. Direct verification of the Weights website shutdown notice was blocked, leaving third-party copies as the primary evidence.
For now, the deal amounts to a quiet talent acquisition that deepens OpenAI’s voice bench—and leaves the lab’s safety-first posture on public voice cloning intact. Unless the companies step forward, the Weights acquisition will remain an unconfirmed footnote in the larger voice-AI story.