Vercel adds DeepSeek search and page fetching to AI Gateway Vercel adds DeepSeek search and page fetching to AI Gateway

Retrieval becomes a separate choice

The integration gives an application two distinct operations. Search finds candidate pages and returns their titles, addresses and summaries. Fetch takes a page address and returns its contents as Markdown. The model can therefore discover a source first, then request the underlying text. Vercel also describes combining these tools with Exa or Parallel search results. A page found by another search provider can become input to DeepSeek Fetch. That separation makes the discovery service, page reader and answering model independent choices within one application.

The bill follows tool calls

A question is not the billing unit. Consider an illustrative workload of ten thousand questions, each producing one successful search and two successful page fetches. That produces thirty thousand tool requests and a $22.50 retrieval bill at the announced rate. Search alone accounts for $7.50; the two fetched pages add $15. These figures describe an assumed workload, not a measured benchmark. A system that reads five pages for each question makes a different number of calls even when it serves the same number of users. The arithmetic separates retrieval volume from audience size.

Result limits also shape context

The documentation allows between one and fifty search results, with ten as the default. Changing that count does not change the price of an individual search request. It can change how much material the application subsequently reads. Host-model input and output tokens remain a separate charge, including tool results supplied to the model. Developer-specified search options take precedence over model arguments. That gives the application an explicit place to define retrieval limits rather than leaving every query parameter to generated tool calls. The same result count can still lead to different reading paths.

The implementation preceded the announcement

The November 2 Gateway package release already lists DeepSeek Search and Fetch among its changes. The same release preserves cancellation and timeout error reasons instead of wrapping them as retryable server errors. That distinction affects whether an application interprets a stopped operation as a reason to try again. AI SDK 6.0.321 includes this Gateway dependency, version 3.0.227. The November 6 announcement names that SDK version as its version-six minimum. A package release and a public announcement are separate milestones; the dependency record makes that sequence explicit.

Fallback needs its own accounting

The announcement’s fallback example asks the model to try DeepSeek, Exa and Parallel in sequence after errors. Vercel describes that order and the retry limit as best effort because the model directs the calls. This differs from a deterministic application rule that stops after a fixed number of attempts. For the illustrative workload, adding a second search attempt changes the request count before any page is read. An answer counter alone would hide that change. Search history, fetched-page history and token use explain different parts of the final bill.