AI voice got great. Now the fight is the business model.

Here is the lay of the land. OpenAI sells reasoning-voice infrastructure (GPT-Realtime-2 + translation + transcription) — picks and shovels for agent builders, priced by token. ElevenLabs (and Hume, Cartesia) sells polished, proprietary, rented voices: pay per minute, get the tooling and the safeguards, don't touch the weights. Mistral's Voxtral and friends (Kokoro, Chatterbox, Fish Speech) let you download a near-frontier voice and run it yourself. Sesame and co. bet on consumer apps where the voice is the product, not the component.

The pressure is on the proprietary camp. When an open model like Voxtral runs on a single consumer GPU, ships its weights for free, and — on Mistral's own numbers — beats ElevenLabs on quality, the rented-voice incumbents cannot charge premium rents for median quality. They can only charge for the things open-weight can't easily replicate: polish at the top of the range, voice-cloning safeguards, developer tooling, uptime SLAs, legal compliance around clone misuse. That's a real business — but it's a narrower one than 'we're the only one with a good voice'. The commodity middle is going open, same as it did with text models.

OpenAI's Realtime stack — GPT-Realtime-2, Translate, Whisper — is not really in the same race as ElevenLabs. It is voice-agent infrastructure: a model that can reason mid-conversation, translate live, and transcribe, all in one low-latency stack. The pricing reflects that — you pay by token for reasoning, not by minute for audio. It is building the voice-first agent layer; ElevenLabs is building the best managed TTS. These are adjacent markets with different buyers. An agent builder on OpenAI's stack might still send audio output through ElevenLabs for the voice quality; the two are not necessarily head-to-head.

The 2026 voice market is not one race — it is a split. Capability converges; differentiation moves to interaction model (OpenAI's agent angle), trust (whose voice clone, with what guardrails), and control (rent vs run). The 'best-sounding' question is already a rounding error. The right questions are: does your use case need reasoning mid-call? Do you need to self-host for cost or privacy reasons? Are you building a developer product or a consumer experience? Answer those, and the camp chooses itself. Most buyers will end up in one of two places: paying ElevenLabs for the tooling convenience, or running Voxtral because the quality is close enough and the cost saving is real.

The one structural bet worth watching: consumer voice apps like Sesame are the wildcard. If talking to AI by voice becomes a daily habit for normal people — not just a developer feature — the market geography shifts significantly towards whoever owns the consumer relationship. That is the bet Apple is doubling down on with the rebuilt Siri, and it is the bet Sesame is entirely built on. The developers and the builders will be fine either way; the interesting question is whether the voice app becomes the interface layer that everything else routes through, or whether it stays a feature inside products that own the consumer directly. The answer to that question matters more for ElevenLabs' long-term story than any near-term quality comparison.