Last updated
Dessa AI API: Access, Pricing and Integration Status
Check Dessa AI API availability, pricing and commercial-use terms, then see which integration details need confirmation before building.
Is there a public Dessa AI API?
We have not found public Dessa AI API documentation, an API signup or published pricing in Block's release material checked on 25 October 2026. This website does not currently provide a Dessa endpoint. An application to a research program is not the same as commercial developer access.
You can prepare the surrounding voice experience now, but a working Dessa integration depends on a real interface and suitable permission. Do not write production code against a guessed endpoint or assume the eventual protocol will match another provider's API.
Which release would make a Dessa AI API integration possible?
| What Block releases | What a developer can evaluate |
|---|---|
| A public consumer application | Whether users can use that application; it does not imply an API |
| A hosted API and developer terms | Authentication, transport, session behavior, billing and allowed use |
| Weights with a suitable license | Whether self-hosting is permitted and operationally practical |
| Research-only access | The permitted research scope; do not assume a right to resell access |
Follow the official documentation for the route that actually appears. If a third party later hosts Dessa, check the provider's relationship to the release, its applicable license and its own data handling. A provider advertising the model name is not sufficient proof.
Can the Dessa AI API power an Agent?
A deployable Dessa voice Agent cannot be confirmed today. Without a published Dessa AI API or usable licensed weights, there is no verified integration route. Full-duplex speech alone does not establish tool calling, retrieval, memory or action execution.
If a Dessa AI API becomes available, it could be evaluated as the conversation component of an Agent. A separate application would still need to identify the user, retrieve permitted records, authorize actions and check their results. Whether Dessa can request those actions directly or needs a separate reasoning model remains unverified.
For comparison, GPT-Live's documented design separates spoken conversation from a backend that handles reasoning and tools. That is a useful architecture to evaluate, not a published Dessa specification.
What will a Dessa AI API integration need?
The model is only part of a usable voice service. Before committing to a Dessa AI API integration, get answers to these practical questions:
- Audio transport: supported input/output formats, streaming protocol and browser or server connection method.
- Conversation control: how the service represents interruptions, concurrent listening and speaking, silence, reconnection and session end.
- Application actions: whether tools, external knowledge or a separate reasoning backend are supported; natural speech alone does not perform an order lookup.
- Operating limits: session duration, concurrent sessions, account limits, regions and failure responses.
- Data handling: audio retention, logging, training use, deletion and any required consent.
These are integration requirements to verify, not specifications that Block has already published. A credible first example must run with the documented model version and show both successful conversation and failure recovery.
How much would a Dessa AI API session cost?
No public Dessa AI API price was verified in the checked material. Free access, a subscription price, a per-minute rate and a token rate are all unconfirmed. This site's future service price would also be separate from the underlying model's price.
For a realistic estimate, calculate the cost of a completed session: model or hosting charges, any reasoning/tool calls, audio transport, optional storage, and failed or reconnected sessions. Divide the total by successful sessions or useful conversation minutes. Do not compare one provider's audio-only rate with another provider's entire application cost.
For example, OpenAI's GPT-Live documentation separates duration-based voice billing from backend-model and tool usage. That illustrates why a headline per-minute number is not enough to establish the cheapest finished voice agent. It is not evidence of how Dessa will be priced.
Would the Dessa AI API allow commercial use?
The checked release page does not establish a commercial license. Wait for the terms covering the actual Dessa AI API or weights you intend to use. Check whether they permit your type of hosted service, redistribution, modifications and use of generated audio, and whether attribution or other conditions apply.
“Research preview,” “runs locally” and “free to download” would each answer a different question. None alone establishes unrestricted commercial hosting. If the terms do not answer your intended use, seek clarification from the publisher before offering a paid product.
Prepare a Dessa AI API evaluation before access opens
Write down the user task and build around a replaceable provider boundary. Prepare microphone permission, connection feedback, an explicit end button and useful reconnect/error states. Keep credentials and usage decisions on the server. Preserve a clear separation between the conversation interface and any business action it requests.
Use the following fixed conversation with a mock order service. It is a proposed acceptance test for a future Dessa AI API integration, not a test we have run. Fixture: order A-104 belongs to test user Sam Lee, is unshipped, and initially goes to 12 Oak Street. Begin with a pre-authenticated test session authorized to read and update that order; saying “Sam Lee” does not authenticate the caller. The mock service must never contact a customer or change a real order.
Run the successful case first with operation ID success-1. Save that ID before sending the update and use it to identify the operation in the mock service's records.
| Spoken input or test event | Expected behavior | Pass condition |
|---|---|---|
| “Check order A-104 for Sam Lee. Do not change anything.” | Read the authorized fixture and report its status | Correct order and status; no write call |
| During the answer: “Wait, I said A-104, not A-140.” | Acknowledge the correction and retain A-104 | No lookup or change to A-140; response resumes coherently |
| “Change it to 21 Pine Street. Read that back before saving.” | Read back the complete address and request confirmation | No write before the next explicit confirmation |
| “Yes, save 21 Pine Street.” | Submit one authorized mock update, then report the returned result | Exactly one update; stored address matches; success only after tool success |
The two fault cases below are separate runs, not additional saves after the successful case. Before each, reset the address to 12 Oak Street and clear the mock operation log, then replay the four prompts above. Use timeout-1 for the timeout run and disconnect-1 for the disconnected run. Within a run, any retry must reuse its saved operation ID.
| Fault injected after the confirmed save request | Expected recovery | Pass condition |
|---|---|---|
| The mock service commits the update, but its reply is replaced with a timeout | Say the result is unconfirmed, then retrieve the outcome for timeout-1 | No success claim before reconciliation; stored address is 21 Pine Street and exactly one update is recorded |
| The mock service commits the update, but the client disconnects before receiving the result | Reconnect and retrieve the outcome for disconnect-1 before considering a retry | The saved operation is recovered; stored address is 21 Pine Street and exactly one update is recorded |
Before evaluating a Dessa AI API, set an interruption-response target appropriate to your app and measure it from audio events; do not invent a universal latency threshold. Record model/version, language, device, transport, all attempts and failures. Run the same fixture in each required language with an independently checked translation. A pleasant voice does not compensate for an unauthorized or duplicated action.
This site's first planned service is browser voice conversation. It will only be labeled available after an authorized integration works end to end. Follow release changes and the separate public access guide.
What if I cannot wait for a Dessa AI API?
Use the developer alternatives to choose between a documented hosted voice interface and a local research stack. The right choice depends on whether you need simultaneous conversation, application tools, local execution or a specific language—not just which model sounds best in a short recording.