DeepSeek Opens API for Its Full Suite of Apps: Do Developers Still Need Professional Tools?
English text is being prepared.
Why read this
AI score 64
The article compares the measured differences between the DeepSeek full suite and professional tools in search, scraping, and file parsing, and sets out the priority order for customers' procurement decisions.
Topics
Body · AI translation
Is DeepSeek starting to take business away from AI tool vendors?
Recently, DeepSeek announced the opening of its "all-in-one app API," providing developers with capabilities such as web search, file parsing, and voice interaction all together, using a single API to serve more needs.
Previously, Alibaba Cloud Model Studio provided DeepSeek models and accompanying tools, and Doubao Work also supported access to third-party models such as DeepSeek. DeepSeek packages tools like search and documents and sells them to developers, while other platforms also take its models and pair them with their own services.
Doubao Work, meanwhile, lets users choose models from different vendors within the same product. As models and tools are increasingly sold in bundles, how much demand do developers still have that needs to be handed to specialized tool vendors?
A developer in the payments field told DiSiBo that vendors such as Alibaba Cloud, Tencent Cloud, and Volcano Engine are platforms to begin with, and offering different models and tools is an extension of their existing business. In his view, AI-native vendors like DeepSeek have a different business focus: general model R&D remains the main line, and the surrounding tool ecosystem serves that main line.
1/ DeepSeek sells tools, other platforms sell DeepSeek
On October 17, DeepSeek opened its self-developed Search and Fetch APIs, used for web search and web page scraping, respectively. On November 13, DeepSeek further integrated these tools into model conversations, while also opening capabilities such as file input, optical character recognition (OCR), automatic speech recognition (ASR), and text-to-speech (TTS).
Currently, V4.1 Flash already supports features such as web search, file input, and link reading under the three interfaces Messages, Responses, and Chat Completions.
After developers enable the web search tool, the model can search web pages on its own based on the question and attach citations in its answers; after enabling deep thinking mode, it can also repeatedly search and read web pages during multi-step reasoning. After uploading PDF, Word, PowerPoint documents, or images, the model can also extract content and analyze it.
The execution of tools such as search and web page reading, as well as the related context processing, is completed on DeepSeek's server side, which can reduce developers' integration and orchestration work. Capabilities such as OCR, ASR, and TTS are also provided as standalone APIs for developers to combine as needed.
DeepSeek has also open-sourced a demo client that supports features such as web search, file upload, and history import. The company says developers can basically replicate the main functions of the DeepSeek App using this set of APIs.
In terms of pricing, DeepSeek's web search and web page scraping both cost 5 yuan per thousand calls, while file parsing and OCR both cost 1.5 yuan per thousand pages. Speech recognition and speech synthesis are in a limited-time promotional period, priced at 0.39 yuan per hour and 1.5 yuan per ten thousand characters, respectively.
Model token fees are billed separately. Taking V4.1 Flash as an example, non-cached input is charged 1 to 2 yuan per million tokens, and output is charged 4 to 8 yuan per million tokens, with different rates for peak and off-peak hours. If a task requires repeated searching, web page scraping, and reasoning, the actual cost also depends on the number of calls to each service and token consumption.
(Image source: DeepSeek API documentation)
DeepSeek is not the first to offer models together with tools such as search and files. In March 2025, OpenAI launched the Responses API, with built-in tools such as web search, file search, and computer operation, and also provided an Agents SDK to help developers orchestrate Agent workflows.
In China, platforms such as Alibaba Cloud Model Studio and Volcano Ark also provide multi-model invocation and accompanying tool services. In other words, developers who want to use DeepSeek models do not necessarily have to go through DeepSeek officially.
For example, Alibaba Cloud Model Studio already supports DeepSeek V4.1 Flash, and developers can use the web search tool provided by Model Studio when calling the model.
Taking a Responses API call in the Model Studio North China 2 (Beijing) region as an example, web search is billed at 4 yuan per thousand calls under the agent policy, with model tokens billed separately. This search unit price is lower than DeepSeek's 5 yuan per thousand calls, but the two differ in search strategy, returned content, and invocation method, so it cannot be used to judge which is cheaper for completing the same task.
Doubao Work, which targets end users, is also adding model options. On November 12, Doubao Work updated its third-party model lineup, adding models such as Kimi K3.1 Pro, GLM 5.3, and DeepSeek V4.1 Flash.
Previously, third-party model invocation capabilities had already been opened to team and enterprise subscription customers. Enterprises can use Doubao Work's AI credits to pay invocation fees in a unified way, without having to separately purchase each vendor's model API, and can also set employees' model usage permissions. Unlike Model Studio, which provides APIs to developers, Doubao Work puts model selection into the application users use every day.
(Compiled from public information from DeepSeek, Alibaba Cloud Model Studio, and Doubao Work)
The industry insider also raised another possibility: "The ecosystem of models and tools beyond the foundation model is, in essence, a data collector for real-world scenarios." In his analysis, search requests can reflect what information users are looking for, and OCR needs can reflect the types of documents actually being processed. These demand signals are hard to obtain purely from public corpora.
He believes that understanding demand in real-world scenarios may also be one consideration behind model vendors' expansion into tool services. However, this is an analysis of business motives and cannot be used to determine whether DeepSeek uses customer invocation data for model training.
2/ Whether it can replace professional tools still depends on the specific task
How much development work the "full suite" can save ultimately depends on whether the tools can complete the task.
On November 14, developer "Programmer Yupi" published a hands-on test of the DeepSeek full-suite API, comparing the web search and web scraping features with the previously commonly used professional tool Firecrawl.
Firecrawl mainly provides developers with services such as web search, scraping, and full-site crawling, and also supports dynamic web page rendering and browser interaction.
According to his disclosure, the test used 3 search terms and 4 URLs. In this limited sample, the results returned by DeepSeek's basic search were roughly comparable to Firecrawl's, with faster response times.
However, when encountering web pages that require JavaScript dynamic rendering, DeepSeek only scraped part of the static content, while Firecrawl obtained the chapter directory. Firecrawl did not outperform on all web pages either: on other test pages, it failed to scrape or returned a large amount of irrelevant content.
In terms of file parsing, he uploaded a 35-page PDF containing a large number of images. DeepSeek could answer questions about the main text but missed some information in the mind map; after switching to a standalone OCR interface, the text was recognized, but there were still errors in the hierarchical relationships.
After testing, he chose to leave ordinary Chinese search to DeepSeek, while continuing to use Firecrawl for tasks requiring JavaScript rendering and full-site crawling. This was a choice made based on personal needs and a limited sample and cannot be directly generalized to all application scenarios.
(Image source: Programmer Yupi; the Firecrawl price in the image was converted by the author based on the package)
Speaking about the impact of the "full suite" on professional tool vendors, the aforementioned industry insider believes that when choosing tools, customers may not necessarily consider price first. According to his observations, most customers are still in the validation stage, and the priority of procurement decisions is "feasibility → convenience → low price": first confirm that it works, then compare user experience and cost.
But not all professional services are easily covered by foundation models. He pointed out that services for processing enterprises' private-domain knowledge are rather special. The difficulty lies not only in search and retrieval capabilities but also in access permissions, enterprises' internal standards, and implicit rules within organizations. Such information is inherently outside the scope of publicly indexable content, and no matter how strong a model's search capability is, it may not be able to solve this directly.
(Image source: Firecrawl official website)
Professional tool vendors are also expanding their services. On November 12, the day before DeepSeek released its full suite, Firecrawl launched MCP v3. According to the official introduction, developers can access more than 3,000 tools through a single MCP connection, with services expanding from web search and scraping to enterprise information, scientific research materials, government records, and more. These tools are provided through aggregation in its directory, which does not mean they are all developed by Firecrawl itself.
Developers may not completely abandon professional tools, but some services that previously required separate purchase now have the more convenient option of the "full suite." For DeepSeek, the boundary of the full suite may also be the boundary of its ecosystem: which capabilities it absorbs itself and which it leaves to external tools. For professional AI tool vendors, finding "scenarios that are not easily covered by foundation models" is a difficult but correct thing to do.
AI Trends