On 13 November 2026, DeepSeek opened the main features of its consumer app to developers through the API. Until this change, many teams had to stitch together their own search, file parsing, link fetching, safety review and speech tools. Those pieces can now be turned on in the existing APIs, or called as standalone services.
The model at the centre of this release is V4.1-Flash, which you call as deepseek-flash. It now supports built-in web search, deep thinking search, file input, URL reading and safety checks across the Messages, Responses and Chat Completions APIs. The service runs the tool calls and assembles the context on the server, so your Laravel or Node.js app does not have to execute the search or fetch the page itself.
This is useful for teams who want a lower-cost way to add “search the web,” “read this PDF,” or “summarise this link” to an assistant. It is not a free add-on. Tool use is billed on top of token charges, and public-facing apps have extra duties around safety and labelling. This article explains what was opened, how billing works as of mid-November 2026, and how to adopt the features without surprises.
Key takeaways
- On 13 November 2026, DeepSeek made App-style search, file input, URL reading, safety checks and voice features available through its API, with
deepseek-flashas the model for the built-in tools. - Enable search with
deepseek_web_searchand URL reading withdeepseek_read_url. The model decides when to use them. You cannot force those tools withtool_choice. - Public-facing apps must enable safety checks. Developers still have to handle their own content moderation, any required model registration, and labelling of AI-generated content.
- Tool use is billed in addition to token charges. As of 15 November 2026, official search and web-fetch prices are $0.75 per 1,000 billed searches or fetches.
- Search, Fetch, OCR, speech-to-text and text-to-speech are also available as standalone APIs if you want those steps in your own workflow, not only inside a chat completion.
What you can turn on in a chat request
The official changelog for 13 November 2026 describes five conversation capabilities on deepseek-flash, available in Messages, Responses and Chat Completions.
Web search is the everyday mode. You declare deepseek_web_search and leave thinking disabled. The model searches the web, or searches by image when the user message contains images, when the question needs it, and it answers with citations. One search call can contain several queries. The client does not run the search; the service does.
Deep thinking search uses the same tool, but with thinking enabled. Over several reasoning steps, the model can search, open pages and find text in those pages. Official docs say it makes up to about 20 tool calls before it answers. Opening a page in this mode is part of the search tool, not the same as reading links the user pasted.
URL reading is enabled with deepseek_read_url. The model looks at the last user message, finds links, fetches those pages, and uses them in the answer. Repeated URLs count once. At most the first 10 different URLs in that message are considered, and the model chooses which of them to read. If thinking is enabled and web search is also declared, that round uses deep thinking search and does not read the user’s links separately. Link reading happens when thinking is off, or when only the URL-reading tool is declared.
File input lets you upload PDFs, Office files, images and more. After plain-text extraction or rich-layout parsing, the content is added to the conversation so the model can answer from it, without first building your own parser.
Safety checks review the model input and output and return a refusal marker when a risk is detected. If you provide a service to the public, official guidance says you must enable this parameter, and you still have to meet your own obligations for content review, large-model registration where it applies, and AI labelling. Setting safety_check to {"level": "high"} turns on this basic risk control. A level of low is the default, the same as omitting the parameter.
Declare only the tools you want. To read links without searching, declare only deepseek_read_url. If you set tool_choice to the name of either web tool, the request returns an error. If you set it to none, all of those tools are turned off, including URL reading. Official docs also say these web tools cannot yet be combined with certain other server tools in the same request.
How billing works
There are two layers of cost: tokens for the model, and per-use fees for tools. The figures below are from DeepSeek’s official pricing page as of 15 November 2026. Product prices can change, so check that page again before you finalise a quote.
For deepseek-flash, official token prices are listed per 1 million tokens, with peak and off-peak rates. Off-peak rates are half of peak rates. Peak hours are 01:00 to 04:00 UTC and 06:00 to 10:00 UTC, Monday through Friday, excluding Chinese public holidays. In India that is 6:30 a.m. to 9:30 a.m. IST and 11:30 a.m. to 3:30 p.m. IST on those weekdays. All other hours are off-peak, including weekends and Chinese public holidays in full.
| deepseek-flash token type | Off-peak, per 1M tokens | Peak, per 1M tokens |
|---|---|---|
| Input, cache hit | $0.003 | $0.006 |
| Input, cache miss | $0.15 | $0.3 |
| Output | $0.6 | $1.2 |
The official model version behind deepseek-flash is DeepSeek-V4.1-Flash-1017. Context length is listed as 1 million tokens, and maximum output as 384,000 tokens. Thinking is the default unless you switch it off. Older names such as deepseek-v4-flash are still accepted, but those models have been retired and the requests are served by this Flash version at the Flash price.
Search and fetch have their own table. The same prices apply whether you call the Search and Fetch APIs directly or the model uses them through the web tools:
| Tool or API | Unit | Official price as of 15 November 2026 |
|---|---|---|
| Search (text or image) | Per 1,000 searches | $0.75 |
| Web fetch | Per 1,000 fetches | $0.75 |
| File parsing and OCR API | Per 1,000 pages | $0.225 |
| Text to speech (list price) | Per 10,000 characters | $0.45 |
| Speech to text (list price) | Per hour of audio | $0.116 |
On the same official page, text-to-speech and speech-to-text also showed a limited-time 50% discount, at $0.225 per 10,000 characters and $0.058 per hour of audio. No end date for that discount was published on the page we checked. If speech is part of your budget, confirm whether the reduced rate still applies at the time you go live.
The billing rules are more precise than “every click costs money,” and they matter in production. A web search that returns results is billed for each query, so three queries count as three searches. Failed or empty searches are not billed. For image search, each image counts as one search, and each region the model selects in an image counts as one more; failed or empty image searches are not billed. URL reading is billed per page fetched successfully, including pages later replaced by a not-allowed or unavailable error under content safety. Links that cannot be fetched are not billed. During deep thinking search, a page is billed only when it is fetched from the web; opening the same page again, moving through a page, opening a page whose content already came with the search results, and finding text in a page are free. Some counts in the API response are for display only. Log the billed tool-search, tool-fetch and tool-image-search figures on every request. A deep thinking search that asks several queries and opens several pages can cost more in tool fees than you expect, even when the token bill looks modest.
Standalone APIs for search, files and voice
If you do not want the model to decide when to search, you can call Search and Fetch yourself and then pass the results into a later model call. Official pricing for those direct APIs is the same $0.75 per 1,000 successful searches or fetches. A direct request is billable only when it returns HTTP 200 and produces a result. Failed requests, empty searches and timeouts are not billed.
OCR is available as a standalone API for recognising text in a PDF, Word or PowerPoint file, or in an image, page by page, as plain text or with rich layout. Official pricing lists file parsing and the OCR API at $0.225 per 1,000 pages. Confirm on the current pricing page how conversation file input is billed, so you do not assume a chat upload is free just because you already budgeted tokens.
Speech-to-text transcribes an audio file in one request. Official limits include a maximum size of 8 MiB, a maximum duration of 120 seconds, and support for 16-bit PCM WAV or Ogg Opus. Streaming recognition is not supported on that API. Text-to-speech can read a block of text, or stream audio while the model is still generating, with a maximum of 10,000 billed characters in a request and 30 minutes of audio in one response. Speed control and resuming an interrupted request are not supported. DeepSeek also published an open-source demo chat client. That is a reference, not a production app. If you use it, treat API keys, history imports and content safety as your responsibility.
Safety, refusals and public apps
For a public-facing service, official guidance says you must enable safety checks, and that those checks do not replace your own duties. When high-level review is on, both input and output are checked. If the input does not pass, the request can return an error. If the output does not pass, the response can end with a refusal after the model has already started talking. In a streaming interface, text that has already appeared cannot be pulled back unless your client removes it. Handle refusals in the client, hide the unsafe answer, and show your own refusal message. The response does not explain the specific reason for the refusal.
If your app streams tokens to a browser, plan for a last-moment refusal. If a page fails review during URL reading, you may receive a not-allowed error for that link. Log it, and do not invent a summary of a page the model was not allowed to read. You also remain responsible for any required model registration and for labelling AI-generated content. Those rules differ by country and by sector. If you ship a public assistant in India, confirm what labelling and registration apply. The API flag is not the whole job.
How to add this to a Laravel or Node.js app
The integration is straightforward, but the operational design is what keeps the bill and the risk under control.
- Start with one job, such as answering a question about a PDF, summarising pasted links, or answering a current factual question with citations. Do not turn on every tool in the first sprint.
- Use
deepseek-flashand declare only the tools you need. Store the model name and tool list in configuration. If a feature only needs to read user links, do not also enable search. - Turn thinking on only when the task needs it. Thinking is the default unless you switch it off. Deep thinking search can make several tool calls, each of which may add fees. For simple questions, non-thinking web search is usually cheaper.
- Enable high safety checks before any public traffic. Handle refusals in both streaming and non-streaming clients. Do not display a half-finished answer that later failed review.
- Pass history back unchanged, including the encrypted signatures on pages that were read. Official docs say the page text is not returned in plain form in the tool block. If your gateway strips unknown fields, later turns may lose that context even though the request still succeeds.
- Log tokens and billed tool counts for every request, and set a per-user or per-tenant cap. Watch peak-hour token rates if much of your weekday traffic falls in those IST peak windows.
- Keep secrets out of prompts and uploaded files, and do not treat citations as a guarantee of truth. For answers that can affect money, health, legal advice or a customer decision, keep a human review step.
A reply that already used the web tools cannot be continued as a partial assistant message; official docs say that request returns an error, and you should resend the previous request instead. Build your client with that limit in mind.
What this does not automatically give you
This release makes a rich assistant easier to build, but it does not remove the usual product work. File input is useful for a document the user just uploaded; a large, changing knowledge base is still often better as retrieval from your own store. The figures here are from the official pricing page as of 15 November 2026, including a limited-time speech discount with no published end date, and prices can change. Safety checks are a required starting point for public apps, not a complete compliance programme. And you cannot force a search or a link read with the built-in tools; if you need a search on every request, use the standalone Search API and then call the model.
How we can help at DayDecoder
DayDecoder is a custom software company in Ahmedabad. We build AI agents, chatbots and RAG systems on Laravel and Node.js. If you want help wiring these tools into an existing product, use the contact form. We reply within one business day.
FAQ
When did these API tools become available?
DeepSeek announced them on 13 November 2026, for deepseek-flash across Messages, Responses and Chat Completions.
Do I pay only for tokens?
No. Official pricing charges token use and, separately, tool use. As of 15 November 2026, search and web fetch are $0.75 per 1,000 billed searches or fetches. File parsing and OCR, and the speech APIs, have their own per-page, per-character or per-hour rates.
Must I turn on safety checks?
If your app is public-facing, official guidance says you must enable safety checks. You still remain responsible for content moderation, any required registration, and labelling of AI-generated content.
Can I force the model to search every time?
Not with the built-in web tools. Naming those tools in tool_choice returns an error. Use the standalone Search API if you need a search on every request.
Written by
Prakash Tank
Founder & Full-Stack Architect, DayDecoder
Prakash started as a certified Laravel developer and grew into a full-stack architect who owns a project end to end: requirements, architecture, development, deployment, and support. He designs event-driven systems with...
More about Prakash