Developer Guide
Dessa AI API for Developers: How to Prepare Before It Exists
Search interest in a Dessa API keeps growing — and this month it turned into clicks. Developers are no longer just asking whether Block's speech-only voice model has an API; they are clicking through to find out. So here is the honest starting point, followed by the useful part: Block has announced no Dessa API — no documentation, no keys, no SDK, no pricing per minute. But there is real preparation a developer can do today, because the shape of a full-duplex speech API is different from the text APIs most teams know, and the teams that understand that difference will be ready first if Block ever opens the door.
The status, without the spin
Dessa was announced on October 23, 2026 as a research preview. Early access goes first to a small group of AI safety researchers, and everyone else can join the waitlist on Block's official page. Our dedicated status article, Dessa AI API: what Block has (and hasn't) announced, tracks every official word on developer access, and the short version has not changed: nothing about an API has been announced. Anyone selling you a "Dessa API key" today is not Block. This article is about something different — what to build, learn and decide now so an announcement, if it comes, finds you ready instead of starting from zero.
Why a Dessa API would not look like a text API
Most developers build voice features as a pipeline: speech-to-text, then a text model, then text-to-speech. Dessa deletes that pipeline. As we explain in how Dessa AI works, it takes audio in and produces audio out directly, listening and speaking at the same time. An API for a model like that would almost certainly be a streaming audio connection, not a request that returns a paragraph. Your app would send a live microphone stream and receive a live voice stream back — with interruptions, overlaps and emotional tone travelling in both directions. That single difference changes your client code, your latency budget and your testing.
Preparation 1: get your audio pipeline streaming-ready
Whatever provider eventually supplies full-duplex speech — Dessa or a rival available today — the client-side work is the same. Capture microphone audio continuously instead of in push-to-talk chunks. Handle echo cancellation and background noise, because a model that hears everything hears your television too. Build playback that can be interrupted mid-sentence, because a full-duplex partner will change its mind when you cut in — and your users will cut in. If your current voice feature waits for silence, transcribes, waits again and then speaks, treat that turn-taking logic as the thing most likely to be rewritten.
Preparation 2: design for conversation, not commands
Block's demos show Dessa backchannelling with "yeah" and "got it", laughing mid-sentence, finishing a speaker's sentence and adjusting its volume when asked — a behaviour Block says nobody explicitly taught it. Interfaces built for that kind of partner need new decisions. When should your app let the model interrupt a user, and when must it never? How do you show, on screen, that the system is listening while it speaks? What does your product do when the model laughs, sighs or sounds frustrated — because a speech-native model produces tone as output, not just words? Writing these rules now, as product policy rather than code, costs nothing and will save arguments later.
Preparation 3: plan the safety layer yourself
There is a reason safety researchers get Dessa first, a rollout we examined in is Dessa AI safe? A voice that sounds human raises impersonation, disclosure and consent questions that no API provider can fully answer for you. If your product ever speaks to customers, decide in advance: will callers always be told they are talking to an AI? What voices may it never imitate? What conversations get recorded, and who can listen? Which requests — payments, account changes, medical or legal topics — must hand off to a human? Block owns Square and Cash App, so expect any future developer offering to arrive with rules attached; having your own policy ready makes adopting them straightforward.
Preparation 4: know what local inference would change
Block says Dessa's inference runs faster than real time locally on a MacBook Pro with an M4 Max chip or better, after training on a single NVIDIA GB200 NVL72 system. That is a demonstrated capability, not a promised product form. Still, it is worth modelling both futures. A cloud API means per-minute costs, network latency and audio leaving the device. A local model means hardware requirements, larger downloads and audio that can stay on the machine. Your architecture should be able to say which future your product prefers — and why — before pricing (which Block has not announced) forces the question.
What not to do yet
Do not build a business dependency on an unannounced API. Do not promise customers "Dessa-powered" anything. Do not buy domains, keys or "early developer access" from anyone but Block. And do not treat today's available realtime voice APIs and Dessa as interchangeable: rivals such as OpenAI's realtime offering are built on text-trained models and are available now, while Dessa's speech-only approach remains unproven under independent testing — no benchmarks, parameter count or evaluation has been published. The honest developer strategy is boring and strong: build a streaming audio client, write your conversation and safety policies, keep your provider layer swappable, and watch the Dessa AI Tracker for the day the status actually changes.
Sources: Block's official Dessa page (block.xyz/dessa) for the model, demos and local-inference claims; DessaAI Daily Search Console data, November 2026, for the developer-search interest described above.
Frequently asked questions
Is there a Dessa AI API for developers?
No. Block has announced no Dessa API, SDK, documentation or developer pricing. Dessa is a research preview, early access goes first to AI safety researchers, and the public can join the waitlist at block.xyz/dessa.
How would a Dessa API differ from a normal voice API?
Dessa is speech-to-speech with no text step, and it listens while it speaks. A developer interface would most likely be a live streaming audio connection rather than a text request-and-response call — but Block has announced no details, so treat that as informed preparation, not a specification.
What can developers do now to prepare for Dessa?
Build a streaming, interruptible audio pipeline; write conversation rules (when the AI may interrupt, how listening is shown); set safety and disclosure policies for a human-sounding voice; and keep your voice provider layer swappable so an unannounced API is never a dependency.
🚀 Get Dessa launch alerts
Dessa is still a research preview. Join the free DessaAI Daily list and we will email you the day access opens — plus the week's biggest Dessa stories, once a week. No spam, no hype.
Form not loading? Subscribe on the signup page →