DeepSeek / Official web chat
DeepSeek Search: find sources and check the answer
To research a changing fact in DeepSeek’s official chat, switch on Search before sending your question. Ask for the original source, then open it and check the exact claim yourself. In our three September 2026 first-answer tests, Search found useful official pages—but one answer also made a confident claim that its cited page did not support.
ChinaAI is independent of DeepSeek. This guide covers its consumer web chat, not an API search endpoint or the model running behind the chat interface.
By China AI Chat Team · Web interface and official sources checked September 23, 2026 · Three original consumer-chat tasks on September 23, 2026 · Our research method
Choose Search, DeepThink or both
In the official web chat we used, Search and DeepThink appeared as separate controls below the message field. Select Search when your answer depends on a current source, such as an updated policy, release note or price. Select DeepThink when the task needs a worked-through answer. On September 23, 2026, our session let us select both controls for one source-based calculation. That is an observation about this interface and session, not a promise of better results or identical availability for every account.
| Your job | Useful starting setting | Still check |
|---|---|---|
| Rewrite text you already supplied | Neither control is necessary for that request. | Every required fact and format constraint. |
| Find a changing fact | Search on; request a direct primary-source link. | Whether the linked page actually states the claim. |
| Work through a self-contained logic or code problem | Try DeepThink if available; Search is useful only when outside facts are needed. | Arithmetic, assumptions or executable tests. |
| Calculate from a current published table | Search and DeepThink can be selected together in the web session we tested. | The table’s date, units, conditions and your calculation. |
DeepSeek’s August 2025 V3.1 release note described the DeepThink button as a Think/Non-Think toggle for that release. Its current API thinking guide separately documents developer request parameters and an API default. Neither document identifies which model served our September 2026 consumer-chat answers. Do not equate the button with one fixed R1 release or copy an API default into the app.
How to verify a DeepSeek Search answer
- Start at the official web chat. Open a new conversation, select Search under the message field, and check that it remains selected before sending. Select DeepThink as well only if you want to try it for a reasoning task.
- State the exact claim and source you need. Name the organization and page type if you know them. For a product setting, ask for the vendor’s current documentation rather than a generic web summary.
- Ask for direct links, not only a numbered marker. Send the question. A marker shows what the answer cites; it does not certify that the cited page contains the sentence beside it.
- Open the page and find the precise sentence, row or footnote. Check model name, date, region, plan and units. A page on the right domain can still be the wrong page or an older release.
- Mark each important claim as supported, partly supported or unsupported. If the cited page does not support the answer, treat that claim as unverified. For a consequential decision, check another current authoritative source or contact the provider.
Copy and adapt this prompt for current documentation. It asks for evidence, but the verification step remains yours:
Using only [organization]’s current official documentation, answer: [specific question]. Give a direct link for every decision-changing claim. If a claim is absent from the linked page, say you cannot verify it. Separate the consumer app from API behavior, and include the page’s update date when available.
For example, if the answer says a model ID will soon be replaced, open the exact cited pricing table or release note and look for the retirement date and routing rule. A model name appearing elsewhere on the same site does not support that extra prediction.
Three first answers we checked against DeepSeek’s own pages
Test conditions: September 23, 2026; signed-in official consumer web chat; three English questions, each in a fresh chat; one first answer per question; no regeneration or corrective follow-up. Search was selected in all three. DeepThink was unselected for the first two and selected for the third. The chat UI did not expose the serving model or plan.
How we added the third task and checked citation links
We added the third task after observing that both controls could be selected, and fixed its question and checks before sending it. We inspected every distinct page cited to support the answer, using its current canonical page when a citation had tracking parameters.
| Question | What the first answer got right | What checking the source changed |
|---|---|---|
| Which API IDs serve which model versions? | It mapped deepseek-flash to DeepSeek-V4.1-Flash and deepseek-v4-pro to DeepSeek-V4-Pro-0813, matching the current official table. | It also claimed that V4 Pro would retire and reroute on September 14. Its cited pricing pages did not say that. More decisively, DeepSeek’s September 10 change log said the V4 Pro API service would continue after that date with unchanged billing. The answer’s extra claim contradicted that current official update. |
| What is the API thinking default and off-switch? | It correctly reported that the current API guide says thinking is enabled by default, and that OpenAI-format requests can set {"thinking":{"type":"disabled"}}. | It said the guide’s Python example shows the disabled setting. The displayed official example actually uses enabled. Its claim that the consumer app defaults to thinking off relied on a 2025 release note; that source does not establish the default for every current session. |
| Can Search and DeepThink work in one source-based calculation? | With both controls selected, the interface displayed a thinking trace, web search and one opened official page. It used the current off-peak deepseek-flash rates and calculated $0.30 input + $0.60 output = $0.90 for the exact hypothetical token counts we gave. | The linked pricing table supported both rates and the arithmetic checked out. This is one observed workflow, not a reliability study or a consumer-chat price. |
The first answer’s retirement claim is a useful warning: correct citations on two rows did not validate a third statement in the same answer. The cited Chinese table had a footnote about legacy Flash names routing to V4.1-Flash; it did not say the listed deepseek-v4-pro ID would retire. DeepSeek’s September 10 change log, which the answer did not cite for that claim, instead states that V4 Pro API service continues after September 14. We cannot determine actual backend routing from this chat test, but the current official documentation contradicts the answer’s prediction.
Original protocol, first-answer text and UI screenshots: download our evidence records (ZIP). The archive excludes account credentials, private chats and API keys. We show these small tests so you can repeat the method, not to assign DeepSeek an overall accuracy score.
Try the source-based calculation yourself
The third task is a compact way to practice the full workflow. We asked the official chat:
Using only DeepSeek’s current official Models & Pricing documentation, calculate the off-peak API cost for 2 million cache-miss input tokens plus 1 million output tokens on deepseek-flash. Show the two line items, total in USD, and a link to the exact table. If you cannot check the current rate, say so. This is an API calculation, not a price for the consumer chat app.
At our review time, the official table gave $0.15 per 1M off-peak cache-miss input tokens and $0.60 per 1M off-peak output tokens for deepseek-flash. So the check is (2 × $0.15) + (1 × $0.60) = $0.90. If the table changes, recalculate from the live table rather than copying our dated result. The actual bill depends on usage, cache status and the applicable time window; see our DeepSeek pricing guide and calculator for those distinctions.
Use the same pattern for release notes, support policies or documentation you rely on: ask one answerable question, name an authoritative source, open the link, and compare each claim with the exact source line. If the answer offers a feature or limit absent from the source, label it unverified before acting on it.
What this guide does and does not establish
These are three first answers in one signed-in web session, not a benchmark. Search results, interface controls, account availability and provider pages can change. We did not test the mobile app, every region, all subjects, paid tiers, repeated runs or whether one mode improves accuracy. The visible controls and traces do not identify the serving model. A source link is a starting point for checking, not proof that the answer used that source correctly.
What if Search fails while ordinary chat works?
Check the separate Search Service status and the selected control before retrying. Our DeepSeek Search troubleshooting steps explain how to narrow the problem and when to use the original source directly.
Can Search and DeepThink be selected together?
Yes in the official web session we checked on September 23, 2026. Our combined-mode calculation showed a thinking trace and source retrieval. If your interface differs, use the controls you can see and verify the final answer the same way.
Does the DeepThink button mean I am using DeepSeek-R1?
The button alone does not establish a particular model release. The 2025 DeepSeek release note explained it as a mode toggle for V3.1 at that time; the app did not name the model serving our test. For releases, API identifiers and downloadable weights, use the model guide.
Is this the same as adding web search to the DeepSeek API?
No. This guide tested Search inside DeepSeek’s official consumer web chat. The API guide covers developer requests and their documented features; do not assume that the chat site’s search control is an API parameter.
For a basic product overview and official access, start at DeepSeek AI. For uploaded documents, see our PDF reading and verification test. Avoid entering confidential material into a chat service without reviewing its privacy terms and your own rules. Report a correction with the claim, current source and date checked. We revisit this guide if the controls or cited documentation change.