DeepSeek adds search, OCR and speech tools to its API
The Chinese AI provider launched a suite of integrated tools for developers, but the move introduces specific technical limits and data privacy considerations.
The short version
- New tools include web search, URL fetching, OCR, and speech recognition (ASR) and synthesis (TTS).
- Speech recognition is capped at 120 seconds per request and does not support streaming.
- File parsing can miss charts and diagrams, while rich layout recognition crops figures into the text.
- All data processed through these tools is subject to Chinese state-security laws.
- Tool usage is billed separately from standard model tokens, with search and fetch costing $0.75 per 1,000 calls.
Why it matters
For developers, these updates provide a convenient way to build complex agents with live web access and voice capabilities in one place. However, the consolidation means all data—including uploaded files and search queries—travels through a single legal jurisdiction.
New capabilities for AI agents
DeepSeek recently updated its API to include a suite of integrated tools designed for developers building complex AI agents. These new features allow models to perform web searches, read specific URLs, and process uploaded documents or audio files. By consolidating these functions, teams can build retrieval-augmented generation (RAG) applications more easily without managing multiple separate providers.
The update includes standalone services for optical character recognition (OCR) and speech synthesis (TTS). Developers can now send web links or search queries directly to the model, which can then decide when to run these tools to find information. The model will provide citations for its answers when the thinking feature is disabled, while it can perform multi-step reasoning with thinking enabled.
The chat features for these tools run on the deepseek-flash model across various API types. This means developers can now integrate live search and voice capabilities into their products using a single vendor. However, the convenience of this consolidation comes with specific technical constraints and costs that differ from standard token billing.
Technical limits and parsing gaps
While the new tools offer broad functionality, they come with notable technical limitations. For instance, the speech recognition service only accepts files up to 120 seconds long and does not support streaming, meaning a client must wait for a full transcript before proceeding. Additionally, for stereo audio files, the system only recognizes the left channel, which may impact the quality of call recordings where speakers are separated.
Document parsing also has specific behaviors regarding visual elements. The text recognition mode extracts only text and ignores charts or diagrams, while the rich layout mode attempts to crop figures and place them in the text. DeepSeek warns that these figures are detected automatically, and there is no guarantee that every image will be successfully captured or included in the model's context.
The standalone OCR service provides text page by page along with position coordinates on a grid. It supports various file types like PDF and PowerPoint, but the entire process of uploading and recognition must finish within 180 seconds. Furthermore, if a document's context becomes too long, the API will return an error rather than silently cutting off the content.
Pricing and tool behavior
DeepSeek bills for these new tools separately from the standard input and output tokens used by the model. Web search and web fetch services both cost $0.75 per 1,000 calls. File parsing and OCR are priced at $0.225 per 1,000 pages, while speech synthesis costs $0.45 per 10,000 characters. Speech recognition is priced at $0.116 per hour of audio, with a temporary discount halving that price.
The search engine behind the tool uses a rerank score to determine the relevance and authority of results. It aims to filter out SEO spam and low-quality AI-generated text from its index. However, independent testing has shown that authority scores can behave at the host level rather than the individual page level, which may influence which results appear at the top of a query.
When the model uses the fetch tool, it presents itself as a Firefox browser from a Hong Kong cloud network. It does not currently use a DeepSeek identifier in these requests. Because the search and fetch loops run on DeepSeek's servers, developers see the results only after the turn is finished, rather than approving each individual call as it happens.
Legal and security implications
Because DeepSeek is a Hangzhou-based company, all data processed through its API is subject to Chinese state-security laws. These laws can require the company to cooperate with national intelligence efforts or provide technical support for security investigations. This legal framework applies regardless of what the company's specific privacy policy states about data handling or user rights.
DeepSeek's privacy policy notes that it stores personal data in the People's Republic of China and may use inputs to train its models. While users have an opt-out right for training, the policy also states that data may be shared with authorities when required by law. There is currently no independent security audit available for the new search, OCR, or speech tools to verify how data is handled in practice.
Developers are encouraged to be cautious with sensitive information when using these hosted services. The documentation describes these tools as hosted services with no self-hosted option, meaning the legal obligations remain with the provider. To limit exposure, users should consider stripping personal data from files before sending them to the OCR or speech recognition services.
What to watch
- Whether DeepSeek provides a self-hosted option for these tools to bypass cloud data paths.
- How the rerank scores for search results evolve as more generative AI content appears online.
- Any updates to the speech recognition limits, specifically regarding streaming support.
Try it
Test the search tool with public web pages to see how it handles citations before sending any private documents to the API.
Sources and links
- OfficialDeepSeek's Nov. 13 release api-docs.deepseek.com
- Officialnormalized to a 0-to-1,000 grid api-docs.deepseek.com
- Officiallast updated Feb. 10, 2026 cdn.deepseek.com
- Paperopen-weight OCR model in October 2025 arxiv.org
- ReportingTech Times: the article we sourced
Liked it? Say why
Add a line for other readers.
No account needed. We never show your email. Comments with links, ads or abuse are held and not shown. Your comment shows after a quick check.
Listen to the AI Wire headlines
Listen to the AI WireListen to our daily AI shows every morning: on your commute, in the shower, wherever you are
Listen to the daily showsFirst reported by Tech Times · Read the original. Written by the SI Agent News Desk from the reporting above.Written with AI from the source above. Every number is checked against the source. Spot a mistake? Corrections.
