TL;DR
What is Block's Dessa?
Dessa is a speech-to-speech AI model from a small research lab at Block, the company behind Square and Cash App. Block announced it on Thursday, October 23, 2026. Block says it was trained "from scratch on nothing but human conversation" (no text and no pretrained models), that it listens and talks at the same time, and that it "runs locally on a MacBook Pro M4 Max or better." Dessa isn't publicly available. There's no price and no release date, and Block is giving early access only to "a small group of AI safety researchers." For now, it's a preview of where phone-answering AI could go.
What happened
Block published a Dessa page with more than a dozen recorded conversations on Thursday, October 23, 2026. Dessa lead Ragavan Thurairatnam and Block cofounder Jack Dorsey announced it on X the same day. The Verge covered it that afternoon, and TechCrunch followed on Friday, October 24.
Block says every clip on its page "is a live recording of a person talking to Dessa" and that "nobody steers it." The lab says it's "a (very) small AI research lab at Block, founded by former members of Dessa, the Toronto deep learning startup," and that it reports directly to Dorsey.
What Dessa does
All of these are Block's claims, from its Dessa page:
- Audio in, audio out. Dessa never converts your speech to text and back. Block argues that text throws away tone, sarcasm, laughter and sighs.
- Full-duplex conversation. It listens and talks at the same time, so it can cut in, finish your sentence or "backchannel" with a "yeah!", "got it!" or "ooh" while you're still talking.
- Faster than real time. Block says inference "runs faster than real time, so there's no awkward wait."
- Runs on a laptop. Block says it "runs locally on a MacBook Pro M4 Max or better." Thurairatnam's post says "consumer GPUs."
- Learned from far fewer words. Block says Dessa learned language "from hundreds of times fewer words" than text-based models. TechCrunch notes that whatever it knows came from conversation data.
- Modest training hardware. Block says it trained Dessa on "only one NVIDIA GB200 NVL72," a single rack-scale AI computer.
Who can use it, and what it costs
| Status as of October 26, 2026 | |
|---|---|
| Public access | None. TechCrunch: "no public access to Dessa at the moment" |
| Early access | "A small group of AI safety researchers," by contacting the lab through Block's Dessa page |
| Price | Not announced |
| Release date | Not announced |
| In Square or Cash App? | Block hasn't said |
| Hardware Block cites | MacBook Pro M4 Max or better |
What reviewers heard
The Verge's Jay Peters found the conversations "occasionally a little awkward and uncanny." In one of Block's own clips, Dessa mistakes the speaker for "Ronaldo." He also pointed to a clip where Dessa chats with OpenAI's GPT-Live-1 and comes across as the more natural speaker of the two. TechCrunch's Ivan Mehta noted that Dessa joins other full-duplex voice systems like OpenAI's GPT-Live. He added that rivals such as ElevenLabs v4 and Fish Audio Drama 3 use voice tags to set emotion, and that it's "not clear" whether Dessa's emotion can be steered the same way.
What it means for small businesses
If your business lives on the phone, with takeout orders, salon bookings or "are you open Sunday?" calls, this is the kind of model you'll be offered in a year or two. TechCrunch points out that "Square already uses voice AI to handle ordering," and that a model running on a device rather than in the cloud could mean lower cost. If you can't wait, one recruiter has already built a phone agent that books into Calendly with tools on sale today.
Three things are worth thinking about now:
- Natural voices change what callers expect. A voice that laughs and says "got it" mid-sentence feels friendlier, and that also makes it harder for callers to tell they're talking to AI. Plan to tell customers when they're speaking with an AI.
- Local can mean private. A model that runs on your own hardware could keep call audio off someone else's servers. Block hasn't said how Dessa would be deployed in products, so this is still a "could."
- Realistic AI voices help scammers too. The FTC warns that if a call "sounds like your boss (asking for bank account numbers)," you're more likely to act. Block itself says it wants safety researchers to "find the risks we haven't anticipated."
For ideas you can use now, browse our customer service AI hacks.
The jargon, in one line each
| Term | What it means |
|---|---|
| Speech-to-speech model | An AI that takes audio in and gives audio out, with no text step in between |
| Full-duplex | Both sides can talk and listen at the same time, like a real phone call |
| Backchannel | Little "mm-hm" and "yeah" sounds that show you're listening |
| Pretrained model | A model someone else already trained, which a team builds on instead of starting from zero |
| Running locally | The AI runs on your own computer instead of a company's servers |
| Voice tags | Labels like "excited" or "whisper" that tell some voice AIs how to sound |
| NVIDIA GB200 NVL72 | A rack-scale NVIDIA system built for training and running large AI models |
How it compares
| Dessa (Block's description) | Typical voice AI (as Block and TechCrunch describe it) | |
|---|---|---|
| How speech is handled | Audio in, audio out | Speech is often turned into text first, then back into a voice |
| Turn-taking | Listens and talks at once | Many wait for you to finish. Some, like OpenAI's GPT-Live, are also full-duplex |
| Emotion | Hears and speaks it natively, per Block. "Not clear" whether it can be steered, per TechCrunch | Some tools, like ElevenLabs v4 and Fish Audio Drama 3, use voice tags |
| Where it runs | Locally on a MacBook Pro M4 Max or better, per Block | Varies by product |
| Can you use it? | No, only early access for safety researchers | Many voice AI tools are on sale today |
X reactions
Jack Dorsey introduced Dessa in Block's own words, and the lab's lead went further with a big comparison.
Our take
The clips are fun to listen to. A model that gasps "oh my god" under its breath during an anniversary surprise is a long way from the robot that reads you your account balance. Training it only on speech, on a single rack, is a genuinely interesting research result.
But "a GPT3 moment for speech" is a big claim to make about a model nobody outside Block can talk to. Every clip so far was picked by the team that built it. We'll believe the hype once ordinary people get to have awkward, noisy, real conversations with it.
For small businesses, the useful move this week has nothing to do with Dessa. Voices this good are coming from many directions, so set up a call-back rule for anything involving money before a convincing voice asks your staff for a bank detail.
— Artificially Nerdy
What to do this week
- If you run a phone-heavy business: listen to two or three clips on Block's Dessa page, then picture your regulars on the other end. Would they want that, or would it feel like a trick?
- If you're shopping for an AI phone answerer now: start with AI support tools that answer calls and chats or AI voice tools you can use today, and ask the vendor two questions. Does it turn speech into text in the middle? Where is call audio processed and stored?
- If anyone on your team handles payments: agree that any phone request to change bank details or send money gets a call back on a number you already have. The FTC recommends verifying "using a phone number you know is theirs."
- If you run a family business: consider the FCC's suggestion of a "safe word" only your family knows, for calls that claim to be an emergency.
- If you're an AI safety researcher: Block is asking researchers to get in touch through the Dessa page for early access.
FAQ
Can I use Block's Dessa?
Not yet. There's no public access. Block is offering early access only to a small group of AI safety researchers.
How much does Dessa cost?
Block hasn't announced a price or a release date.
What does full-duplex mean in voice AI?
The AI can listen and talk at the same time, so it can react while you're still speaking, the way people do on a phone call.
Does Dessa run on a laptop?
Block says it "runs locally on a MacBook Pro M4 Max or better." You can't download it, though, since it isn't publicly available.
Will Dessa be in Square?
Block hasn't said. TechCrunch notes that Square already uses voice AI for ordering, but there's no announced plan to put Dessa into Square.
Who built Dessa?
A small research lab at Block led by Ragavan Thurairatnam, founded by former members of the Toronto deep learning startup Dessa. It reports to Jack Dorsey.
Sources
- Block: Dessa — Block's official page with the model description, live recordings, the hardware claims and the safety researcher call.
- TechCrunch: Block's new voice model sounds eerily human — Ivan Mehta's October 24, 2026, report on availability, rival models and the Square angle.
- The Verge: Jack Dorsey's Block made an AI model that talks to you like an awkward new friend — Jay Peters' October 23, 2026, impressions of Block's recorded clips.
- jack on X and Ragavan Thurairatnam on X — the October 23, 2026, announcement posts.
- FTC: Fighting back against harmful voice cloning — the FTC's advice on calls that sound like your boss or a family member.
- FCC: "Grandparent" scams get more sophisticated — the FCC's "safe word" suggestion for family emergency calls.
