AI voice in 2026: who makes what

The map. Four distinct camps: OpenAI's reasoning-capable Realtime voice models (GPT-Realtime-2, Translate, Whisper) launched May 2026; ElevenLabs and peers (Hume, Cartesia) for polished proprietary quality; Mistral's Voxtral (March 2026, free weights) and Kokoro, Chatterbox, Fish Speech for open-weight models you run yourself; and consumer apps like Sesame (launched May 2026, iPhone, four agents, from Oculus founders). OpenAI's GPT-Realtime-2 brings reasoning into a voice model — it thinks mid-conversation, not just speaks. Voxtral runs on a single consumer GPU with free weights, and Mistral says it beats ElevenLabs on quality.

The shift. Voice is becoming a primary interface to AI, not an optional mode — Apple rebuilt Siri around it, OpenAI bundled reasoning into it, consumer startups are betting whole products on it. The commodity middle of the TTS market is going open-weight: Voxtral's quality-to-cost ratio puts real pressure on mid-tier proprietary pricing. ElevenLabs and peers survive on tooling, safeguards and developer experience — not on owning the only good voice.

The catch. Open-weight means you run your own infrastructure — GPU, uptime, updates. For teams without that, the proprietary services still win on convenience, with cloning safeguards and developer tooling the open camp can't yet match. And OpenAI's reasoning-voice stack is a developer API priced per use — excellent for agents, but a different product from a managed text-to-speech service, so it's the wrong tool (and the wrong bill) for simple narration. The 'best-sounding model' question barely matters now; the right questions are reasoning, control and who you trust to run the voice.