A research agent can ask a search service for documents, inspect what comes back, and decide which passages to send to its model. It can also hand the entire job to a model endpoint that runs its own search. The answer may look the same. The application no longer makes the same decisions between the question and the answer.
That is the consequential change in DeepSeek’s November 13, 2026 application API release. Our October assessment examined standalone Search and Fetch and the Flash-1017 model update. This release lets DeepSeek run retrieval and assemble conversation context inside the model request. My assessment is that the managed route is a sensible candidate for public research with flexible budgets. It needs a different admission test for workflows that require pre-approval of sources, exact external call limits, or replayable evidence. DeepSeek changelog
A compatible interface can carry different controls
DeepSeek documents these capabilities for deepseek-flash across Messages, Responses, and Chat Completions. The hosted search tool can find web or image results; enabling thinking lets the model perform a longer search-and-read sequence. A separate URL-reading tool can act on links in a user message. File input adds extracted or parsed material to the context. The release also lists standalone OCR and speech APIs and basic model content review. These are documented interfaces, not results from a live integration. November release
The most useful section of the chatbot guide is the one on ignored parameters. In the documented managed mode, requests can carry fields that the service does not honor. For example, the Responses API list includes max_tool_calls, parallel_tool_calls, and store; the Chat Completions list includes seed. The guide says its output limit applies to each inference step rather than to the aggregate round. The parameter exceptions can also be triggered by high safety checking or automatic titling. Ignored parameters by protocol
An application must distinguish a requested value from an enforced limit. If a workflow may perform at most one external read before review, sending a field named max_tool_calls is insufficient on the documented managed route. A local timer or disconnect can bound how long the user waits; neither establishes how much remote work has already happened, whether it was cancelled, or what will be billed. Where the grant requires exact call-by-call authorization, keep the tool gate under application control or leave the workflow out of the managed lane.
Likewise, an ignored store field tells you that this field cannot establish the promised handling for that route. It does not establish what the service actually retains. Check the relevant provider terms and deployment evidence before admitting a class of documents. A compatibility label does not resolve that question.
Separate what a user can cite from what an auditor can replay
The developer guide shows search and fetch result blocks, citation markers, and usage fields. It instructs the client to carry the returned blocks forward without modification. Its streaming example includes a search result with encrypted_content, as well as visible fields such as a title and URL. The example is documentation, not a demonstration that every source or action is observable in every protocol. Chatbot response and streaming examples
That gives a user useful navigation. It does not necessarily give a reviewer the exact content that conditioned the answer. A link may change, a result may include only an excerpt, and an opaque state object may be useful to the provider without being inspectable by the client. Treat three records as different things: the provider’s carry-forward state, the evidence a reviewer can read, and the action or claim for which that evidence was accepted.
Consider a support agent comparing two public release notes. It can cite the right vendor and the wrong historical page version. A reviewer needs to see the particular passage as it existed when the answer was composed, with the document date and the effective date distinguished. When your policy allows it, retain the relevant visible excerpt or an independently captured artifact and the retrieval time. When the capture is missing, mark the answer unverified for that use even if the link renders correctly.
Model content review also belongs in a different layer from source permission. A refusal marker can support a user-facing content decision. It does not prove the caller was entitled to submit a file or that an external redirect was in scope. A legitimate question can contain a confidential identifier. Decide that boundary before enabling provider tools, not after reading the generated answer.
Measure the delegated work, including tails
DeepSeek’s public pricing table lists $0.75 per 1,000 searches and the same rate per 1,000 fetches; it applies per-use charges to direct and model-invoked tools. Model tokens and file parsing are separate quantities. At those listed rates, four billable searches and six billable fetches cost $0.0075 before tokens, parsing, application overhead, and review. That arithmetic is hypothetical. The actual number of calls is the variable here. Pricing and deduction rules
A managed route can save engineering effort where an application would otherwise maintain ranking, retries, and a crawler. It can also move costs into extra searches, repeated fetches, generated tokens, and claims a reviewer must chase. Compare cost per accepted, adequately supported answer, counting unsuccessful attempts and human checks in the numerator. Report the median and tail separately. A slow deep-search path triggered by a few ambiguous requests can dominate the customer experience without moving the mean much.
A proposed adoption experiment
Choose public, permitted material for the first pilot. Build paired questions from stable pages, changed pages, missing pages, and conflicting sources. Include a release-note question whose answer has changed and a question that should finish without retrieval. Hold the requested model and scoring rubric fixed. Archive the available inputs and result metadata for both routes, respecting permitted retention.
| Comparison | Evidence to capture | Failure that changes the decision |
|---|---|---|
| Application-managed retrieval versus hosted search | Source selections, claim support, total cost | Correct-looking answer cannot be independently checked |
| Hosted search with thinking off versus on | Call counts, fetched pages, elapsed time | More steps add cost without recovering the missing evidence |
| Single-page task versus conflicting-page task | Cited passage and effective date | A recent page is treated as if it governed earlier events |
| Bounded wait versus ordinary completion | Provider usage after timeout, retry outcomes | Retrying duplicates cost or silently loses evidence |
Use the managed route where it stays inside the admitted data boundary, delivers a useful evidence record, and meets the measured tail budget. Keep the explicit route where the system must select, inspect, or authorize each read. This is a decision about the workflow, not a ranking of model quality.