Comparison
Dessa vs OpenAI's realtime voice models
OpenAI sells two kinds of realtime voice models through its API, and GPT-Live-1 is the full-duplex one. You can use it today. Block announced Dessa on 2026-10-23, and it is not available yet. This page compares the two and explains where GPT-Realtime-2.1 fits.
Updated
| Compared | DessaBlock | GPT-Live-1OpenAI |
|---|---|---|
| Availability | Announced 2026-10-23. Early access for a small group of AI safety researchers only. No price. | Generally available in OpenAI's API since 2026-09-10. $0.05 per minute of voice session, billed per second. The backend model is billed separately. |
| Open weights | None published. | No. API only. |
| License | None published. | Closed model, used under OpenAI's API terms. |
| Training data | Block says: recordings of human conversation only, no text. Amount not published. | Not published. |
| Text backbone | None, according to Block. | Not published. Reasoning and tool use go to a separate backend model that you configure. |
| Model size | Not published. | Not published. |
| Latency | No figure. Block says inference runs faster than real time. | 0.798 seconds of turn-taking latency in OpenAI's own tests (1.41 seconds for GPT-Realtime-2.1). |
| Languages | Not published. | OpenAI lists voices for English, Portuguese, German, French, Hindi, Japanese and Korean. |
| Hardware | Block says it runs locally on a MacBook Pro M4 Max or better. | Runs in OpenAI's cloud only. |
| Hosting options | None today. | OpenAI's Live API, over WebRTC, WebSocket or SIP for phone calls. |
Our take
For anything you need to ship now, GPT-Live-1 is the practical choice of the two. It is full duplex, documented, priced and reachable from browsers and phones.
Dessa is a research result, not a product yet. Speech-only training and laptop-class inference could matter for cost and privacy, if Block releases the model and its claims hold up.
Either way, keep the voice model swappable in your design. We plan to host Dessa if its license allows it, and other full-duplex models either way.
Which OpenAI model to compare
OpenAI's Realtime API runs the GPT-Realtime models, most recently GPT-Realtime-2.1, released on 2026-07-06. They are speech-to-speech, but they work in turns: voice activity detection decides when the user has finished. GPT-Realtime-2.1 calls functions itself and lets you set its reasoning effort, where more effort can mean more latency.
OpenAI's Live API runs GPT-Live-1, which became generally available on 2026-09-10. OpenAI says it listens and speaks at the same time and handles pauses, interruptions and backchannels. That makes it the closest OpenAI match for Dessa, and Block used it in two of its own demo clips.
Sources: OpenAI, OpenAI, OpenAI, OpenAI Developer Community, Block
Two different designs
GPT-Live-1 splits the work. The voice model runs the conversation, and anything that needs reasoning, search or tools goes to a backend model: one that OpenAI runs through its Responses API, or your own agent. OpenAI notes that interrupting the voice does not cancel backend work by itself. While a session runs, your app can also add instructions, facts, or lines the model may say aloud.
Block describes Dessa as a single model that learned everything it knows from human speech. TechCrunch notes that this differs from the usual setup, where a speech model is paired with a large language model for knowledge. Block has not said whether or how Dessa could call tools or pass work to another model.
Sources: OpenAI Developer Community, OpenAI, Block, TechCrunch
Connecting and paying
GPT-Live-1 is a working product. You can connect over WebRTC from a browser, over WebSocket from a server, or over SIP for phone calls. Outbound calls over SIP work once OpenAI enables them for your organization. Voice time costs $0.05 per minute, billed per second, and the backend model is billed at its usual rates.
GPT-Realtime-2.1 is priced per token instead: $32 per million audio input tokens and $64 per million audio output tokens, with cheaper cached input. It also works over WebRTC and WebSocket, and accepts incoming phone calls over SIP.
Dessa has no API, price or connection method yet.
Sources: OpenAI, OpenAI, OpenAI, OpenAI, OpenAI Developer Community, OpenAI, Block
Voices and languages
OpenAI's GPT-Live docs list voices for English, Portuguese, German, French, Hindi, Japanese and Korean. Since 2026-10-16, API customers can also create custom voices from a text description.
Block says Dessa chooses its own voice and personality in each conversation. It has not said which languages Dessa speaks or whether developers will be able to pick a voice.
Latency claims
OpenAI reports 0.798 seconds of turn-taking latency for GPT-Live-1 in its own tests, against 1.41 seconds for GPT-Realtime-2.1. Block gives no number for Dessa. It says only that inference runs faster than real time on a high-end MacBook Pro.
Neither claim has been checked independently, and no test has put the two models side by side under the same conditions. If latency matters for your product, measure it end to end on your own network.
Sources: OpenAI Developer Community, Block
Cloud or local
The biggest difference may be where the model runs. GPT-Live-1 runs only in OpenAI's cloud. Session storage is off by default; if you turn it on, OpenAI keeps session recordings for 30 days.
Block says Dessa runs on a MacBook Pro M4 Max or better. If Block ever releases it in a form that runs on your own hardware, apps could keep conversations on the device. Block has not said whether it will.
Where each one stands
GPT-Live-1 keeps changing. On 2026-10-15 OpenAI added a Decisions integration, so an app can act while the user is still speaking or react to what the model says. On 2026-10-16 it added nine voices for German, French, Hindi, Japanese and Korean.
Block's Dessa page reads the same on 2026-10-31 as on launch day. The download and launch links are still marked as coming soon, and Block has not given dates.
Sources: OpenAI, Block, Internet Archive
Full-duplex voice API
Build on full duplex with our hosted API.
Pick a plan, and get Dessa news in the same step if you want it.
FAQ
Questions
Is GPT-Live-1 the same as the OpenAI Realtime API?
No. GPT-Live-1 runs on OpenAI's Live API and is full duplex. The Realtime API runs GPT-Realtime models, such as GPT-Realtime-2.1, which detect the end of a turn with voice activity detection.
How much does GPT-Live-1 cost?
OpenAI charges $0.05 per minute of voice session, billed per second, plus the usual rates for the backend model and any tools.
Did Dessa really talk to GPT-Live-1?
Block's Dessa page has two clips of Dessa talking with GPT-Live-1. Block notes that both voices are AI-generated and that the developer of GPT-Live-1 is not affiliated with Block and has not endorsed Dessa.
Can I use Dessa instead of GPT-Live-1 today?
No. Dessa has no API, weights or license yet, and only a small group of AI safety researchers has early access.
Sources
- Changelog, OpenAI
- Voice activity detection (VAD), OpenAI
- Introducing GPT-Live-1 in the API, OpenAI Developer Community, 2026-09-10
- Dessa - the Conversation Model trained only on speech, Block
- Block's Dessa page, archived copy from launch day, Internet Archive, 2026-10-23
- Managing GPT-Live sessions, OpenAI
- Block's new voice model sounds eerily human, TechCrunch, 2026-10-24
- WebRTC, OpenAI
- WebSockets, OpenAI
- Telephony and SIP, OpenAI
- GPT-Live 1 Model, OpenAI
- GPT-Realtime-2.1 Model, OpenAI
Independent site. Not affiliated with or endorsed by Block, Inc.