DeepSeek Opens Its App Features as APIs: Search, Files, URLs, Voice and Safety Checks

DeepSeek opened the features behind its chat app to developers on November 13, 2026. Its API change log calls it "the full suite of application APIs": web search, file input, URL reading, model safety, speech recognition (ASR) and speech synthesis (TTS). For the chat-style features you add a parameter to an API you already call. The service then handles tool calls and context assembly for you.

The important detail is the model. The change log says deepseek-flash supports these features in the Messages, Responses and Chat Completions APIs. That means a developer who already uses the Chat Completions format can get a cited web answer without building a search pipeline. This is the latest step in a month of moves: DeepSeek opened its Search API on October 17, and billing for server-side search tools started on November 1.

TL;DR: what DeepSeek released and what it costs

What can developers now switch on?

DeepSeek's change log lists five chat-level features. Each is a parameter, not a new endpoint.

Web search. With thinking disabled and the deepseek_web_search tool enabled, the model runs web or image searches as the question requires and answers with citations.

Deep thinking search. With thinking enabled and the same tool, the model searches the web and reads links on its own across several reasoning steps. This is the agent-style mode. It will use more tokens and more search calls, so cost control matters (see below).

File input. You upload PDFs, Office files, images and more. DeepSeek extracts plain text or parses the rich layout, adds the result to the conversation context, and the model answers from it.

URL reading. With the deepseek_read_url tool enabled, the model detects links in the user message, fetches the pages and answers with them in context.

Model safety. With the safety check parameter enabled, DeepSeek reviews model input and output and returns a refusal marker when it finds a risk. The company is blunt about the limits: "if you provide a service to the public, enable this parameter and fulfill obligations such as content safety review, large model registration and AI labeling." That language reflects Chinese regulatory duties. It is a basic filter, not a compliance program.

Speech waveform turning into text lines, representing the new DeepSeek ASR and TTS application APIs

Beyond those, DeepSeek sells the underlying pieces on their own: Search and Fetch, OCR, ASR and TTS. The change log also says it built and open-sourced a demo chat client on these APIs, which can import your DeepSeek chat history.

What does it cost?

The Models and Pricing page lists the numbers as of November 13.

Model tokens are billed separately. For deepseek-flash the page lists $0.30 per million cache-miss input tokens at peak and $0.15 off-peak, and $1.20 per million output tokens at peak and $0.60 off-peak. Off-peak is half the peak rate. Peak hours are 01:00 to 04:00 and 06:00 to 10:00 UTC, Monday to Friday, excluding Chinese public holidays.

One important detail: search and fetch are billed per use, whether you call the Search and Fetch APIs directly or the model uses them through tools. A deep thinking search can run many searches for one question. Budget for the search line, not just tokens.

On the October launch, TechNode reported a price of RMB 5 per 1,000 successful requests, with failed requests, empty results and timeouts not charged. The current dollar price on DeepSeek's page is $0.75 per 1,000. We did not verify the no-charge-for-failures rule against the current docs, so check the Search and Fetch guide before you rely on it.

Why this matters: DeepSeek is moving up the stack

Until recently DeepSeek mainly sold models. This release says the layer around the model, search, file handling, safety and voice, is now also a product. That puts DeepSeek closer to what other labs already offer as built-in tools, and it follows its own harness work. Our recent coverage shows the pattern:

• DeepSeek Harness reached 1.54 million daily users on the official API on November 4, with a plugin ecosystem and a planned marketplace.
• The Super-agent design describes a Router plus isolated Workers with a Watchdog.
• DeepSeek V4.1 Flash 1017 shipped as MIT-licensed open weights on October 17, the same day the search API opened.
• The desktop harness for macOS and Windows put the agent on users' machines.

For a builder, the practical gain is fewer moving parts. You can drop a separate search vendor, a PDF parser and a speech vendor if DeepSeek's quality is enough for your task, and you keep one bill and one key. The trade-off is that you accept DeepSeek's index, parsers and filters, and you cannot swap them without changing your code.

The caveats: reliability, data and policy

Reliability. The service had a five-hour outage on October 15, and DeepSeek refunded a week of usage as credits. A single-vendor stack for search, files and voice concentrates that risk. Keep a fallback.

Data. Files you upload, links you ask the model to fetch and audio you send all go to DeepSeek's servers. The change log does not describe retention for these new features in the text we read. Read the privacy terms before sending customer documents or recordings.

Policy. US civilian agencies were ordered to remove DeepSeek and products built on it. If you sell to government customers, check your contract terms before adding these APIs to a product.

Safety. The safety parameter is a basic filter. DeepSeek itself says public services still have legal duties beyond it.

Evaluation. The company posted no benchmark for search quality, citation accuracy or ASR and TTS accuracy in the change log. Treat answer quality as unmeasured until you test it on your own questions.

How would you try it?

We have not run these calls ourselves, and the change log names the tools and parameters but not full request bodies in the text we read. Start from DeepSeek's own guides.

1. Create or reuse an API key and keep the base URL https://api.deepseek.com (OpenAI format) or https://api.deepseek.com/anthropic (Anthropic format).
2. Set the model to deepseek-flash. The older names deepseek-v4-flash and deepseek-v4-flash-vision-exp still work, but those models are retired and requests are served by V4.1 Flash 1017 at the Flash price.
3. Enable the deepseek_web_search tool for a cited answer. Compare thinking off and thinking on, and log the number of searches per answer.
4. Enable deepseek_read_url and test with a page you know well, so you can check the answer against the source.
5. Send a file you can verify. Ask for a quote with a page number and check it.
6. If your service is public, enable the safety parameter, and plan your own review process.
7. Set a daily spend cap before you test deep thinking search.

The deepseek-flash model also has a 1M-token context window and up to 384K output tokens, according to the pricing page, so a large file plus search results can fit in one call. Large inputs cost tokens too.

How does it compare with building it yourself?

We have not tested search quality across vendors, so this table compares only what each approach involves, not answer quality.

For a broader view of how DeepSeek's models stack up on price and capability, see our frontier models comparison and our small models comparison. Third-party tools already built on Flash include OpenCode's free Flash tier.

What is confirmed and what is open

Confirmed from DeepSeek's own pages: the November 13 release, the five chat features, the deepseek-flash model, the three API formats, the standalone Search and Fetch, OCR, ASR and TTS APIs, the prices above, and the open-sourced demo client.

Confirmed by an outlet for the earlier step: the October 17 Search and Fetch launch, per TechNode.

Not public in the text we read: search quality numbers, ASR and TTS language coverage, data retention terms, rate limits for the new tools, and when the half-price speech offer ends.

The takeaway is simple. DeepSeek turned its app into an API surface. For cost-sensitive builders that is worth a test this week, with a spend cap and a fallback in place.

Related reading

• DeepSeek Harness hits 1.54 million daily users
• DeepSeek V4.1 Flash 1017: open weights, benchmarks and pricing
• DeepSeek Super-agent architecture
• DeepSeek's five-hour API outage and credit refund
• DHS orders DeepSeek removed from civilian agencies
• Small models compared: Haiku 5.5, GPT-6 Luna, DeepSeek V4.1 Flash

Details are accurate as of November 13, 2026. DeepSeek can change prices and limits; check the official change log and pricing page.