AI Gateway adds DeepSeek Search and Fetch tools

Vercel has launched native DeepSeek Search and DeepSeek Fetch tools inside AI Gateway, providing web search and Markdown extraction through a single API key. As reported in Vercel Blog's report, these helpers work with any model supporting tool calling without requiring an independent DeepSeek account. Search returns page titles, URLs, and summaries, while Fetch retrieves full content as Markdown.

This technical brief covers feature capabilities, pricing rules, architecture benefits, and code integration patterns for software engineers and AI developers deploying Vercel AI SDK versions 6 and 7 in production environments.

What shipped in AI Gateway

The release introduces two specialized helper tools directly into Vercel AI Gateway: DeepSeek Search and DeepSeek Fetch. DeepSeek Search queries the live web to find relevant pages, returning structured results that contain page titles, URLs, and summaries. DeepSeek Fetch takes a specific target URL and retrieves the entire webpage content formatted as clean Markdown. By returning standard Markdown rather than raw HTML, the tool enables models to read documentation, articles, and external websites without complex custom scrapers or local parsing pipelines.

Both tools work across any model that supports tool calling, avoiding proprietary vendor lock-in. Developers can declare both helpers within the tools configuration object of the streamText function, allowing models like deepseek/deepseek-v4.1-flash-1017 to search the web and inspect retrieved content in a single operation. Because these tools integrate directly into the gateway abstraction layer, teams can switch underlying model providers at any time while retaining identical search and fetch tool definitions across their entire software stack.

What improved in AI Gateway

The primary operational improvement is the removal of multi-vendor account friction. In earlier workflows, integrating web search and content retrieval often required establishing distinct third-party provider accounts, tracking isolated payment ledgers, and building custom scrapers. Vercel AI Gateway now centralizes this entire workflow under a single AI Gateway API key, eliminating separate DeepSeek account requirements. The gateway handles routing, monitors usage and costs across models, and manages execution retries from a unified endpoint, removing operational complexity from production systems.

Pricing and failure protection have also been updated to provide predictable operating expenses. DeepSeek Search and DeepSeek Fetch each cost $0.75 per 1,000 billable requests, while underlying model token usage is billed separately. Crucially, failed requests, connection timeouts, and empty search results incur zero tool charges. These built-in billing protections ensure that unexpected network drops, missing web resources, or rate limits do not generate unexpected charges for engineering teams operating automated agent loops at high volume.

What you gain from AI Gateway

Using DeepSeek tools inside AI Gateway provides architectural consolidation and streamlined governance. Developers no longer need to provision, distribute, or rotate separate DeepSeek credentials across deployment environments. Instead, applications authenticate every inference call, search query, and content fetch using one AI Gateway key, or choose to bring their own provider keys. Centralized usage tracking and cost accounting across models and tools give engineering leads comprehensive visibility into expenses, preventing sudden billing surprises while simplifying infrastructure management.

Teams also gain resilient search fallback capabilities that prevent agent failures during upstream service disruptions. Applications can configure secondary tools such as gateway.tools.exaSearch and gateway.tools.parallelSearch alongside DeepSeek Search, instructing models like zai/glm-5.3-flash to query alternatives if the primary tool fails. Because DeepSeek Fetch is decoupled from search providers, it can retrieve Markdown from URLs returned by any configured search engine, ensuring autonomous workflows complete tasks without being tethered to a single search index.

How to get AI Gateway

Accessing the new search and fetch helpers requires upgrading project dependencies to supported Vercel AI SDK releases. Vercel has published these tools in AI SDK version 6.0.321 and later on the v6 release line, as well as version 7.0.160 and later on v7. Developers can install the latest package release through their preferred package manager and subsequently configure their required gateway authentication token within their local shell environment or production cloud hosting settings:

pnpm add ai@latest

export AI_GATEWAY_API_KEY="your-api-key"

model: 'deepseek/deepseek-v4.1-flash-1017'

model: 'zai/glm-5.3-flash'

In application code, developers import gateway and streamText from the ai package, declaring deepseek_search and deepseek_fetch within the tools object. For example, deepseekSearch accepts a count parameter to set result limits, while deepseekFetch processes target URLs. When configuring multi-provider fallback chains, developers can also import exaSearch and parallelSearch alongside required toolChoice settings and system prompts, allowing the model to search external sources, fetch page content, and stream structured summaries directly to standard output.

What to watch after AI Gateway

Engineering teams implementing search fallbacks should note that provider fallback sequencing is model-directed rather than strictly enforced by gateway routing rules. Because the model interprets error signals and determines which tool to call next, execution ordering and retry limits operate on a best-effort basis. Developers must write explicit system instructions that guide models to attempt alternative search providers upon encountering errors, verifying through execution logs that models follow the planned fallback order rather than getting stuck in repetitive loops.

Teams must also monitor context window token consumption alongside flat tool request fees. Although failed calls, timeouts, and empty results incur no tool charges, successful queries and Markdown fetch operations inject external text directly into the model context window. Ingesting large Markdown documents increases token usage, which is billed separately from the $0.75 per 1,000 tool request rate. Developers should tune search result counts and inspect document sizes to keep inference costs aligned with planned application budgets.