Setting model to deepseek-flash does not turn search on. On 13 November 2026 DeepSeek opened the tools behind its chat app on the API. The id deepseek-flash is DeepSeek-V4.1-Flash-1017. Legacy ids deepseek-v4-flash and deepseek-v4-flash-vision-exp still route here and bill at the Flash price. Context is 1,000,000 tokens. Maximum output is 384,000. Thinking is the default.
This page is dated 16 November 2026 from the 13 November changelog, the web tools guide, the Search and Fetch guide, the safety guide, and the pricing page. Three days ago we covered Gemini 4 Carbon usage. Here the fields that change the request and the bill are tools, thinking, safety_check, and usage.server_tool_use.
What opened on 13 November
The same model accepts the tools on three endpoints. Messages is POST https://api.deepseek.com/anthropic/v1/messages with x-api-key. Responses is POST https://api.deepseek.com/responses. Chat Completions is POST https://api.deepseek.com/chat/completions. The last two send Authorization: Bearer. Search, Fetch, OCR, ASR, and TTS are also standalone calls. Image search is not on POST /v1/tools/search. It runs only inside deepseek_web_search.
| Capability | JSON that turns it on | What comes back |
|---|---|---|
| Web search | deepseek_web_search, thinking disabled | The model searches when it chooses and cites |
| Deep thinking search | deepseek_web_search, thinking enabled | search, open_page, find_in_page, about 20 calls |
| URL reading | deepseek_read_url | Links in the last user message, at most the first 10 unique URLs |
| Safety | safety_check.level = high | Input and output review. DeepSeek requires it for a public service |
| Files | Files API file_id | PDF, Office, and images parsed before the chat turn |
Request JSON on three endpoints
Messages names each tool twice, type and name. Responses and Chat send type only. tool_choice cannot name these tools. That returns 400, deepseek_web_search cannot be forced. tool_choice none turns every declared web tool off, including URL reading. On Responses and Chat, required is treated as auto.
Do not send allowed_domains, blocked_domains, or user_location. Do not declare the same tool twice. Do not mix in another server tool. The error says other tools cannot be used together with deepseek_web_search, deepseek_read_url, safety_check, or title yet. On Responses, a web_search tool, and text.format other than text, return 400. On Chat, web_search_options and response_format other than text return 400. Structured JSON and these web tools do not share one request. Put your own schema on a later call. See what structured output is.
Thinking disabled is the citation search. Thinking enabled is deep thinking search. If both deepseek_web_search and thinking are on, the round does not also run deepseek_read_url. Links are read when thinking is disabled, or when only deepseek_read_url is declared. Messages and Chat set thinking.type to disabled or enabled. The Responses sample for deep thinking search sets reasoning.effort to high.
{
"model": "deepseek-flash",
"thinking": {
"type": "disabled"
},
"safety_check": {
"level": "high"
},
"tools": [
{
"type": "deepseek_web_search",
"name": "deepseek_web_search"
},
{
"type": "deepseek_read_url",
"name": "deepseek_read_url"
}
],
"messages": [
{
"role": "user",
"content": "What changed in the DeepSeek API on 13 November 2026?"
}
]
}Responses, with thinking search:
{
"model": "deepseek-flash",
"reasoning": {
"effort": "high"
},
"safety_check": {
"level": "high"
},
"tools": [
{
"type": "deepseek_web_search"
},
{
"type": "deepseek_read_url"
}
],
"input": "Which official page states the height of Tokyo Tower?"
}Blocks, kind, and citations
On Messages, each call is a server_tool_use block plus a result. Match them with tool_use_id. In a stream both blocks arrive whole on content_block_start. Read final usage from message_delta. The usage on message_start covers only the first inference step.
| input.action | name | kind | Result |
|---|---|---|---|
search | deepseek_web_search | search, or deep_search while thinking | web_search_tool_result |
open_page | deepseek_web_search | deep_search | web_fetch_tool_result. data is an empty string. No signature |
find_in_page | deepseek_web_search | deep_search | Same shape. title is null. Not billed |
read_link | deepseek_read_url | read_link | web_fetch_tool_result plus signature |
Image search sets kind to image_search or deep_image_search. queries is an empty array. A result may add image_url for display only. Citations on Messages sit on text blocks as web_search_result_location. cited_text is the answer text of that block, not a quote from the page. encrypted_index is an empty string. One URL per block. In a stream they arrive as citations_delta.
Responses puts the call in a web_search_call item with kind and action. The answer is output_text with url_citation annotations. start_index and end_index count Unicode code points. Chat puts calls in message.server_tool_calls. Each step has reasoning_content and calls. message.content has no citation markers. Citations are message.annotations. message.reasoning_content is only the last step.
A reply that used these tools cannot be continued. The API returns 400 and tells you to resend the previous request. On a later turn, pass the assistant content back unchanged, including kind and signature. Search snippets and earlier thinking are not replayed into context. Only the answer text is. A page read from a link comes back from signature. For read_link, data is always an empty string. Drop the signature and the request still succeeds, but the page is gone. Failed fetch codes are url_not_accessible, url_not_allowed, and unavailable. url_not_allowed also covers a page that failed the safety review.
One Responses search item:
{
"type": "web_search_call",
"id": "srvtoolu_s_00_3f2a9c",
"status": "completed",
"kind": "search",
"action": {
"type": "search",
"queries": [
"DeepSeek API 13 November 2026"
],
"sources": [
{
"type": "url",
"url": "https://example.com/news",
"title": "Change log",
"snippet": "Web search and URL reading are now API tools.",
"published_at": "2026-11-13T00:00:00+00:00"
}
]
}
}Standalone Search and Fetch JSON
When your code, not the model, decides to search, call the raw JSON APIs. Search is POST https://api.deepseek.com/v1/tools/search. query is required, 1 to 128 characters, not all whitespace. count is 1 to 50 and defaults to 10. ori_question is optional, also 1 to 128 characters. localization is an optional IP. need_content defaults to false. Unknown fields are ignored. Operators that pass through: site:, -site:, -, -"...", "...", lang:, after:, and before:.
{
"query": "DeepSeek site:api-docs.deepseek.com after:2026-11-01",
"count": 10,
"need_content": false
}Fetch is POST https://api.deepseek.com/v1/tools/fetch. The only field is url, at most 8,192 characters, HTTP or HTTPS, with a host and no username or password. The body is title and content as Markdown. A missing title is an empty string.
{
"url": "https://api-docs.deepseek.com/updates"
}| Field | What to store |
|---|---|
rerank_score | 0 to 3. 1.5 and above is the high-quality band in the guide |
authority | 0 to 4. −1 means unknown. 3 and above is highly authoritative |
snippet / summary | Short text for a UI, longer text for a model |
content | Empty unless need_content is true, or when the page has no content |
date | RFC 3339 UTC, or null |
file_id, OCR, and voice JSON
File input is two steps. POST https://api.deepseek.com/files with purpose before file. text_recognition drops charts. rich_layout_recognition crops figures into the text and spends image tokens. The file object comes back only after parsing. id looks like file-api-.... A disconnect does not cancel a parse that then succeeds: the file is stored and billed. You cannot download the original or the parse. Chat Completions references it like this:
{
"role": "user",
"content": [
{
"type": "file",
"file_id": "file-api-xxxxxxxxxxxxxxxx"
},
{
"type": "text",
"text": "Summarize the revenue chart."
}
]
}Messages needs the header anthropic-beta: files-api-2025-04-14 and a document block whose source.type is file. Responses uses input_file and file_id. A parsed document cannot be sent as an image block. You cannot paste the bytes or a file URL in the chat request. One request may cite at most 20 documents and 8 MiB of their text. The full parse stays in context on later turns. Too long returns 400. Nothing is truncated quietly.
Limits: PDF, Word, PowerPoint, and common images, detected from the bytes. 100 MiB, or 413. 1,000 pages. 120 seconds to upload, or 408. 120 seconds to parse, or 504. 15 document uploads in progress per account. Password-protected files return 400. Plain text, csv, json, source code, xlsx, xls, epub, and mobi use purpose user_data and are not billed. Those files max out at 32 MiB, and the text at 8 MiB. AZW3 and DRM ebooks are not supported.
Parsing is $0.225 per 1,000 pages, once, only when parsing succeeds. Later chat turns bill input tokens, not pages again. OCR is the same page price on POST /v1/tools/ocr and does not enter the Files API. It is multipart, file plus purpose, not a JSON body. ASR is multipart audio, not JSON. TTS is JSON, POST /v1/tools/tts:
TTS list price is $0.45 per 10,000 characters, and the published limited-time price is $0.225. A Chinese character counts as 2. Any other character counts as 1. Voices are stella (default) and tide. There is no language field. ASR list price is $0.116 per hour of audio, limited-time $0.058, billed by the second, with a fraction rounded up to 1 second. The pricing page does not give an end date for the half-price window.
{
"text": "Hello from DeepSeek.",
"voice_id": "stella",
"format": "wav",
"filter_markdown": true
}safety_check and refusal JSON
For a service the public can reach, set safety_check to {"level":"high"}. low is the same as omitting the field. Any other level is 400. The check does not replace your own review, registration, or labeling duties. Input that fails returns 400 Content Exists Risk. Those input tokens are still billed, at the cache-hit price. After the model finishes, the connection stays open while the output is reviewed. Text already streamed cannot be pulled back. Replace what you showed. Tokens used before the refusal are billed. Review tokens sit inside output tokens and also in usage.output_tokens_details.safety_tokens on Messages and Responses, or usage.completion_tokens_details.safety_tokens on Chat. A blocked image can name a path such as messages.1.content.0, in stop_details.explanation, a refusal content item, or message.refusal. A request with the web tools and no safety_check can still end in a refusal.
| API | Refusal JSON |
|---|---|
| Messages | stop_reason refusal, stop_details.type refusal. Non-streaming drops thinking and answer text, keeps tool blocks and one empty text block |
| Responses | status incomplete, incomplete_details.reason content_filter. output keeps only executed web_search_call items |
| Chat | finish_reason content_filter. content is an empty string. annotations is omitted. server_tool_calls stays, with each step's thinking cleared |
Bill tool_search, not web_search_requests
Bill tokens and tools on separate lines. Flash prices are per 1,000,000 tokens. Peak hours are 01:00–04:00 and 06:00–10:00 UTC, Monday through Friday, excluding Chinese public holidays. Every other hour is off-peak, including weekends and Chinese public holidays for the whole day. Off-peak is half of peak. cache_creation_input_tokens stays 0.
| Window | Cache hit / 1M | Cache miss / 1M | Output / 1M |
|---|---|---|---|
| Off-peak | $0.003 | $0.15 | $0.60 |
| Peak | $0.006 | $0.30 | $1.20 |
Search and web fetch are $0.75 per 1,000 successful uses, whether the model called the tool or you called /v1/tools/search or /v1/tools/fetch. A tool search that returns results is billed per query, so three queries count as three. count does not change the Search API price. HTTP 200 with an empty results array is not billed. Fetch is billed only when page content comes back. A page that safety then replaces with url_not_allowed or unavailable was still fetched, so it is billed. Opening the same page again during deep thinking search is free. find_in_page is free.
Search is limited to 100 requests per second per account. Fetch is 25, plus a per-domain cap. The caps are per account, not per key, and they are separate from the Flash concurrency limit of 2,500. In the web tools guide, one citation search with two queries reports web_search_requests 1 and tool_search 2. Store tool_search, tool_fetch, and tool_image_search. Do not add web_search_requests on top. On Responses and Chat, server_tool_use omits that trio when all three are 0. thinking_tokens appears when thinking is on. safety_tokens appears when safety_check is high. Both are already inside the output token total.
If generation fails after the HTTP status is already 200, the body is an error plus usage. Messages streaming ends on an error event, with no message_delta after it. Check the body. An SDK will not raise on status 200. The published sample for that two-query search. thinking_tokens and safety_tokens were not on this object. They show up on their own when thinking is on, or when safety_check is high, and they are already inside output_tokens:
{
"input_tokens": 1905,
"cache_creation_input_tokens": 0,
"cache_read_input_tokens": 0,
"output_tokens": 72,
"service_tier": "standard",
"server_tool_use": {
"web_search_requests": 1,
"web_fetch_requests": 0,
"tool_search": 2,
"tool_fetch": 0,
"tool_image_search": 0
}
}JSON to verify before launch
- The body has
deepseek-flash. Messages tools include bothtypeandname. Thinking isdisabledwhen you wanted citation search. - A public app sends
safety_check.levelhigh.response_formatandtext.formatstaytexton these calls. - The parser matches
tool_use_idandkind, and keepssignaturefor the next turn. - The bill uses
tool_search,tool_fetch, andtool_image_search, plus cache-miss input, cache-hit input, and output. It does not addweb_search_requests. - A standalone
queryis 1 to 128 characters.countis 1 to 50. Empty results are not a charge.
Check the JSON locally with the JSON validator, JSONPath ($.usage.server_tool_use.tool_search), and JSON Diff against the previous request. Nothing is uploaded.
FAQ
Does thinking change the search?
Yes. thinking.type disabled is one web search with citations, and kind is search. enabled is deep thinking search: search, open_page, and find_in_page, about 20 calls, and kind is deep_search. Thinking is the default. If thinking and deepseek_web_search are both on, that round does not also run deepseek_read_url.
Can tool_choice force deepseek_web_search?
No. Naming the tool returns 400, deepseek_web_search cannot be forced. The model decides. tool_choice none turns the declared web tools off, including URL reading. On Responses and Chat, required is treated as auto.
Should the bill use web_search_requests?
No. That count includes errors and is for display. Bill tool_search, tool_fetch, and tool_image_search. In the web tools guide, one search block with two queries reports web_search_requests 1 and tool_search 2. A search that returns results is billed per query. Empty or failed searches are not billed.
Does POST /v1/tools/search do image search?
No. Image search runs only inside deepseek_web_search. kind is image_search or deep_image_search, and queries is empty. Each image counts as one search, and each region the model selects counts as one more.
What does a public app send for safety?
DeepSeek says a public service enables safety_check with level high. low matches omitting the field. A risky input returns 400 Content Exists Risk, and those input tokens are still billed at the cache-hit price. A refused answer is stop_reason refusal, Responses incomplete with content_filter, or Chat finish_reason content_filter.
Can I continue a reply that searched?
No. A reply that used deepseek_web_search or deepseek_read_url returns 400. Resend the previous request. On the next turn, pass the assistant blocks back unchanged, including kind and signature. Without a valid signature the request still succeeds, but the page is not put back into context.
Summary
The 13 November 2026 tools change the request, the blocks you store, and a second line on the bill. Declare deepseek_web_search and deepseek_read_url on deepseek-flash. Set thinking off for citation search and on for deep thinking search. Public apps set safety_check.level to high. Bill tool_search, not web_search_requests.