# Parallel Search on Vercel AI Gateway: Set Fast Mode to Get the Free Offer

Parallel search is free through January 31, but a model choosing fast mode is not enough. Configure the tool explicitly and keep model-token costs visible.

By Paras Madan | 2026-11-19


Vercel made Parallel web search in fast mode free through January 31, 2027. Outside eve's default integration, you must explicitly set fast mode in the tool configuration to qualify. Letting the model choose fast mode can still produce a billed search.

The [November 18 announcement](https://vercel.com/changelog/parallel-web-search-is-free-on-ai-gateway-and-eve-through-january-31) applies to AI Gateway and eve. It does not make the whole agent free, and it does not cover every Parallel search mode. The practical setup question is where your application chooses the mode, not merely which mode appears in a successful result.

## Put the free mode in configuration

For the AI SDK integration, the relevant configuration is:

```ts
gateway.tools.parallelSearch({ mode: 'fast' })
```

Treat this as a configuration fragment, not a complete application. Import and register the tool in your existing AI SDK request, then verify the actual execution path in a test environment. I have not run this integration or measured its results.

Vercel's [web-search documentation](https://vercel.com/docs/ai-gateway/models-and-providers/web-search) makes the billing distinction explicit: a model-selected fast call is not eligible unless the tool configuration sets it. Without that setting outside eve, basic remains the fallback mode. Other modes remain billable.

I would check every place the application creates its tools. A shared helper may be correct while a background research worker constructs another instance without the setting. Search the code for both the tool factory and any request-level configuration that can override it.

## Check whether eve is using AI Gateway

The announcement says eve's built-in search uses Parallel fast mode automatically for models served through AI Gateway. There is no separate Parallel API key to supply for that route.

However, the [eve built-in-tools guide](https://eve.dev/docs/concepts/built-in-tools) distinguishes Gateway models from direct-provider models. Gateway uses Parallel unless a search provider is selected; direct-provider models use supported native search instead. A provider override changes the route, so “this application uses eve” is not enough to establish eligibility.

Before comparing bills, record the requested model route, search provider and any custom tool definition. If you recently replaced the built-in search with your own implementation, inspect that code rather than assuming the new default applies to it.

For a team with multiple agents, I would put the choice in one reviewed configuration module. Keep an explicit exception for workflows that need another provider or search mode, with a reason and a separate cost allowance.

## Make a failed free search visible

The Gateway documentation says free capacity is shared and can be rate-limited. That error returns to the model without failing the overall request, and Gateway does not automatically retry it as paid search. Model tokens are still charged, and free searches do not appear as billable calls in `gatewayToolCalls`.

The application therefore needs to distinguish a completed answer from a successfully grounded answer. If search fails, the model may still produce text. That text should not quietly inherit a “verified with current sources” label.

My proposed test set includes a normal lookup, a query with no useful evidence and a simulated search failure. Inspect the final answer and user-facing status in each case. Decide whether failure should stop the task, produce an explicitly unverified response or wait for another attempt. Do not introduce a paid fallback without deciding that budget policy first.

## Evaluate shorter excerpts against the actual task

Vercel describes fast mode as returning relevant pages with short excerpts and supports filtering by site or publication date. That is useful for finding evidence, but it does not establish that every answer can be verified from the returned passage alone.

For a migration question, I would start with official documentation and ask whether the result actually contains the version, deadline or exception needed. Keep the source URL with the claim. If the excerpt omits a crucial qualification, fetch or read the source before presenting a firm conclusion.

Use the same small set of real questions when comparing fast mode with a paid alternative. Score whether the answer found the right source and preserved its limitations, not just whether it sounded complete. Faster discovery is valuable only when the remaining verification work does not erase the benefit.

## Budget for the offer ending

The documented regular price is $5 per 1,000 Parallel requests, and billing returns after January 31. At that rate, 10,000 searches would represent $50 in search charges before model-token costs. That is an illustrative calculation, not a forecast of your application's bill.

Record search count during the promotion even when its price is zero. Separate search charges from model usage and any other infrastructure costs, then review the projected total before the offer ends. A temporary discount is a useful evaluation window; it should not become an unnoticed dependency in your pricing model.

I'm Paras. I work at the intersection of AI, engineering and distribution.

I help AI teams turn pricing changes and developer workflows into practical technical stories.

[Book a 30-minute call](https://cal.com/paras-madan/30min).
