DeepSeek removes separate fees for model-triggered web search
The AI company is eliminating extra costs for search and web fetching when triggered by its models, though standalone API calls remain paid.
The short version
- DeepSeek removed separate charges for server-side search and web fetch triggered during model calls on November 18, 2026.
- Users still pay for the input and output tokens used during these interactions under the standard model pricing.
- Standalone Search and Fetch APIs remain billable at 5 yuan per 1,000 successful calls.
- The change applies to web tools and web_search tools in the Responses and Anthropic-compatible APIs.
Why it matters
This change lowers the cost for developers building AI agents that need to browse the web. It simplifies the billing structure for integrated workflows while keeping high-volume, direct API calls as a separate paid service.
The new billing structure
DeepSeek has updated its pricing policy to eliminate specific fees for tools that occur during a model interaction. Starting at noon on November 18, 2026, Beijing time, the company will no longer charge a separate fee for server-side search and web fetch when these actions are triggered by the model. This means that when an AI model decides to look up information or grab a webpage to answer a prompt, the extra tool cost disappears.
The change specifically targets the web tools and web_search tools provided through the Responses API and the Anthropic-compatible API. It is important to note that this does not mean all DeepSeek search services are now free for users. The company is still charging for the model-token usage, which covers the text the AI generates and processes during the interaction.
As reported by pandaily.com, the update is a billing-policy adjustment to existing paths rather than a launch of new features. Developers who rely on these model-side routes will see a simplified cost structure. They will only need to account for the standard tokens used by the model rather than tracking individual tool execution costs.
Model-triggered vs. standalone
A key distinction in this update is the difference between tools run inside a model interaction and standalone requests. When a developer uses a model-side route, the AI determines when to use a tool like text search, image search, or page fetch. In these scenarios, the specific tool fee is removed, but the input and output tokens remain billable under the applicable model pricing.
In contrast, direct client calls to the standalone Search and Fetch APIs remain subject to their original per-request prices. These endpoints, specifically /v1/tools/search and /v1/tools/fetch, are intended for applications that call them directly without an AI model making the decision. These requests are not exempt from fees just because they are part of an automated agent or application workflow.
DeepSeek defines a billable standalone request as an HTTP 200 response that produces a result. For the Search API, this means returning a nonempty results array, while the Fetch API must return actual page content. The company does not charge for failed requests, timeouts, or searches that return empty results. Furthermore, the price per request does not change based on the number of search results returned.
Technical configurations
The official guide provides specific examples of how these routes differ in practice for developers. For instance, the DeepSeek Harness tool uses the model-side route when the web-search provider is set to DeepSeek server-side search or the default Auto setting. This configuration benefits from the removed tool fees because the model is managing the search process.
If a developer chooses to use the DeepSeek Search API instead of the default provider, they are opting into the separately billed client route. This distinction is vital for cost management, as the billing depends on the specific endpoint or provider selected in the integration. The practical check for a developer is which route their code actually hits, rather than how they describe their AI agent's behavior.
Image search also has specific limitations that developers should keep in mind regarding this update. The standalone Search API does not currently support image search, meaning it cannot be used as a standalone billed endpoint for that purpose. Currently, image search is only available through the web tools, which are covered by the pricing change for model-triggered actions.
Impact on developers
This policy change is designed to make it more economical for developers to build complex AI agents that require real-time information. By removing the extra layer of tool fees, DeepSeek allows these agents to perform multiple web fetches or searches without incurring a high volume of individual request charges. This encourages the use of the Responses API for more dynamic, multi-step reasoning tasks.
Developers will need to audit their current integrations to see which routes they are using. If an application is calling the /v1/tools/ endpoints directly, the cost of 5 yuan per 1,000 successful calls will still apply. However, if the application is using the model-side routes where the AI decides to fetch data, the cost structure is now significantly leaner.
The move reflects a broader strategy to simplify the usage of DeepSeek's web tools. By aligning the cost of model-triggered tools with standard token pricing, the company makes it easier for users to predict their expenses. It removes the complexity of tracking separate charges for every time an AI model decides to browse the internet to find a fact.
What to watch
- How this pricing change affects the adoption of DeepSeek for complex AI agent workflows.
- Whether DeepSeek will eventually add image search to the standalone Search API.
Try it
Check your current DeepSeek integration to see if you are using the /v1/tools/ endpoints or the model-side web_search provider.
Sources and links
- Reportingpandaily.com: the article we sourced
First reported by pandaily.com. Written by the SI Agent News Desk from the reporting above.Written with AI from the source above. Every number is checked against the source. Spot a mistake? Corrections.
Liked it? Say why
Add a line for other readers.
No account needed. We never show your email. Comments with links, ads or abuse are held and not shown. Your comment shows after a quick check.