DeepSeek Opens the Full Suite Application API, Including Standalone Speech Recognition and Synthesis Interfaces
DeepSeek announced the opening of its full suite application API. Capabilities such as web search, deep-thinking search, file upload, and link reading can all be called individually through standalone interfaces and billed separately; the speech recognition and speech synthesis interfaces are being opened for the first time. The company says developers can replicate the complete DeepSeek App 1:1, and it has open-sourced a lightweight chat front-end demo that supports importing historical conversations. On pricing, speech recognition is originally 0.78 yuan/hour, now at a 50% discount of 0.39 yuan/hour; speech synthesis is 1.5 yuan per 10,000 characters, and OCR is 1.5 yuan per 1,000 pages; the V4.1-Flash model already supports the relevant conversation extension features under the three types of interfaces: Messages, Responses, and Chat Completions.
Why read this
AI score 73
The original text lists the calling methods and 50%-off pricing for each standalone interface, which can be used to estimate the integration cost of foundational capabilities such as speech and OCR.
Topics
Body · AI translation
Just now, DeepSeek announced that it is officially opening up the API for its full suite of applications.
What does that mean?
Previously, features that everyone found handy in the DeepSeek App—such as web search, deep-thinking search, file upload, and link reading—can now all be used directly by developers through the API, and they don't have to be bundled together; each has its own independent interface, supporting separate calls and separate billing.
Key point: voice capabilities have also been opened up for the first time, including interfaces for speech recognition and speech synthesis.
The official announcement also included a rather interesting line:
Developers can even replicate a complete DeepSeek App 1:1.
Wait, seriously? They're allowing others to copy their own product??
DeepSeek also thoughtfully provided an example themselves, directly open-sourcing a lightweight chat frontend demo that supports importing DeepSeek conversation history.
In other words, developers no longer need to struggle from scratch with basic features like search, file parsing, and speech recognition, and can quickly build a product with functionality close to the DeepSeek App.
Once the news broke, the comments section got quite lively.
One netizen exclaimed:
What do you mean, replicate a complete DeepSeek App 1:1??? The light of AI!
Overview of the full-suite application APIs opened up
Specifically, DeepSeek opened up these capabilities all at once:
There are two changes worth mentioning separately.
First, web search is actually not being opened up for the first time. As early as October 17 this year, DeepSeek had already opened its self-developed Search and Fetch interfaces.
But at the time, developers still had to handle a lot of the tool-calling and model-integration work themselves.
Developers had to write code to string this entire workflow together themselves.
Now it's much simpler.
Developers only need to enable the web search feature for the model; DeepSeek will automatically handle when to search, what to search for, which web pages to read, and how to continue answering based on the search results.
Second is voice—this time there are finally official independent interfaces.
Previously, although the DeepSeek App already supported voice interaction, the publicly available model API did not provide equally complete voice capabilities.
Developers who wanted to build a DeepSeek assistant that can listen and speak usually had to integrate separate speech recognition and speech synthesis services.
Now, both ASR and TTS can be called through DeepSeek's official interfaces.
Moreover, the underlying capabilities mentioned above don't all have to be used as a bundle.
DeepSeek also provides independent interfaces such as Search, Fetch, OCR, ASR, and TTS, which developers can choose as needed and combine freely.
Currently, the V4.1-Flash model already supports the relevant conversation extension features under the three types of interfaces: Messages, Responses, and Chat Completions.
In other words, Teacher D has now handed the entire set of underlying capabilities behind its own App to developers like building blocks.
So the question is:
With such comprehensive features, what about the pricing?
Speech recognition costs less than 0.4 yuan per hour, and DeepSeek is even offering 50% off this time
First, let's look at the tool API price list DeepSeek published this time:
The most eye-catching, of course, is voice.
Speech recognition is originally priced at 0.78 yuan/hour, and is currently directly 50% off, at only 0.39 yuan.
Converted, that's about 0.0065 yuan per minute; recognizing one hour of audio costs less than 0.4 yuan.
Speech synthesis is also 50% off: originally 3 yuan per 10,000 characters, now only 1.5 yuan.
What does that mean?
Compare it with the speech API prices of major foreign vendors:
OpenAI's GPT-Transcribe costs $0.27 (about 1.81 yuan) to recognize one hour of audio.
Even the cheaper GPT-4o mini Transcribe costs $0.18 (about 1.21 yuan).
Google Cloud's standard speech recognition service, calculated at $0.016/minute, costs $0.96 per hour (about 6.43 yuan).
By comparison, DeepSeek's current 0.39 yuan is only about 1/4.6, 1/3.1, and 1/16.5 of the prices of the above three services, respectively.
Text recognition OCR is 1.5 yuan per 1,000 pages, averaging only 0.0015 yuan per page recognized.
Looking at it this way, DeepSeek's pricing approach this time is quite clear.
For developers who want to build applications such as voice assistants, meeting notes, and document processing, at least the cost of calling basic tools has been pushed to a fairly low level.
DeepSeek's existing conversational product is only a relatively basic form.
That remark is also consistent with Teacher D's usual style.
For a long time, compared with putting great effort into operating consumer-facing products, DeepSeek seems more willing to focus its energy on model research and development.
Although the DeepSeek App has already accumulated a large number of consumer users, from the perspective of product form, it is still the familiar chat assistant.
Now, it has simply opened up all its capabilities, and even prepared sample code for replicating its own APP, giving off a vibe of:
I'll keep researching models; as for how to play with applications, I'll leave that for everyone to explore.
How should I put it? This move is pretty clever...
Looking back at the mobile internet era, super apps like WeChat and Douyin almost reshaped the way people socialize, entertain, and even obtain information.
From ChatGPT to DeepSeek, the most familiar interaction method for everyone is still opening a chat box, entering a question, and waiting for an answer.
Right now, no one has the answer.
That being the case, rather than shutting the door and guessing on your own, it's better to lay out all the tools and let developers around the world try together.
What if someone really does come up with the next hit?
It really is Teacher D's big stage—if you have a creative idea, come on up. (doge)