What happened
Vercel’s announcement says DeepSeek Search and DeepSeek Fetch are now available through AI Gateway. Search finds relevant pages and returns titles, URLs, and summaries. Fetch retrieves page content as Markdown for a model to read. The tools can be used with any model that supports tool calling, and both are accessed with one AI Gateway API key; a separate DeepSeek account is not required. Vercel says the same tools can remain in place when a team switches model providers.
The helpers are available in AI SDK 6.0.321 and later on v6, or 7.0.160 and later on v7. Vercel’s example imports gateway and streamText, configures a model, and passes gateway.tools.deepseekSearch({ count: 3 }) and gateway.tools.deepseekFetch() in the tools object. The example then streams the model’s text output. The changelog also shows a setup with DeepSeek Search, Exa Search, and Parallel Search, instructing a model to try alternatives when a search errors and then fetch a result URL.
Vercel describes that fallback as model-directed: provider order and retry limits are best effort. DeepSeek Fetch can retrieve a result URL returned by any of the shown search tools. That separates the search provider from the page-retrieval step, while leaving decisions about retry behavior to the model and its tool instructions. The announced configuration therefore exposes search, retrieval, and fallback candidates in one tool set rather than presenting Fetch as dependent on a single search service.
Each of DeepSeek Search and DeepSeek Fetch costs $0.75 per 1,000 billable requests. Failed requests, timeouts, and empty results have no tool charge. Model tokens are billed separately. The announcement directs developers to the Search and Fetch documentation for configuration options and pricing details. AI Gateway itself is described as a single API for calling models, tracking usage and cost, and configuring routing, retries, and failover; users can authenticate with an AI Gateway key or bring a provider key.
What the archive adds
New Runtime’s earlier coverage on Browserbase Search and Fetch in AI Gateway (2026-10-08) described web retrieval through the gateway’s tool interface. This is a new integration and a new pricing detail on that broader story: the Search and Fetch roles now include DeepSeek tools with a stated per-request charge.
Trend signal
This supports agent protocols become the interoperability layer: the same retrieval interface can be used across model providers, while search and fetch can be combined as distinct tools. It does not establish how reliably agents choose among several exposed search tools; Vercel calls fallback order and retry limits best effort.
Why it matters
Teams often bind retrieval to a model provider or write separate integration code for search and page extraction. Vercel’s setup offers a common gateway surface for both steps and a way to keep them when changing the model, which can reduce the friction of testing different model providers in an existing agent workflow. Fetch accepts URLs from the demonstrated search tools, so a team can vary search providers without changing the retrieval step.
The pricing also gives teams a starting point for estimating tool costs independently of model tokens: each search or fetch request is listed at $0.75 per 1,000 billable requests, with certain unsuccessful outcomes excluded from tool charges. That is useful for budgeting, but it is not a full cost-per-task measure. Builders still need to account for model usage and determine whether retries, relevance, and retrieved content quality fit their workloads. The fallback example is especially worth testing rather than treating as deterministic routing, because the model chooses the sequence and retry behavior is best effort.
Limits and uncertainty
The announcement does not report search relevance, fetch coverage, latency, or comparative performance against other retrieval tools. It does not provide a measured cost per completed agent task; model tokens are billed separately, and real workflows may use multiple searches or fetches. Vercel characterizes fallback ordering and retry limits as best effort, so the documented sequence is not a guarantee of deterministic failover. The changelog also does not describe service-level guarantees or independently verified behavior under production workloads.
Skill to build now
Add a small retrieval test to an existing tool-calling workflow: issue representative queries, inspect the titles and summaries returned by Search, then check whether Fetch produces usable Markdown for the chosen URLs. If you enable fallback, record which provider the model selected and how many tool calls each task used. Compare the observed outcomes and tool charges with your current retrieval path before making it the default.
Reader takeaway
AI Gateway now offers DeepSeek-powered search and page retrieval behind a model-provider-neutral tool interface. The entry point is specific: meet the stated AI SDK version requirement, use a gateway key, and budget $0.75 per 1,000 billable requests for each tool. Treat the fallback example as best-effort behavior that needs workload testing, not as guaranteed routing.