Parallel Fast Web Search Free on AI Gateway and eve

Parallel fast web search is free on AI Gateway and now the default for eve agents.
AI Gateway launches a free policy for Parallel web search: fast mode is free until January 31 of next year, covering any model on the gateway. Developers need no separate Parallel API key—just declare it in a tool call—and zero data retention is enabled by default. Parallel is also set as the default web search for eve agents, with the built-in web_search tool automatically using fast mode when models are served through AI Gateway, skipping account, key, and result-parsing setup. AI SDK users can mount search via gateway.tools.parallelSearch({ mode: 'fast' }) and limit multi-step calls with stepCountIs. Search returns links with brief excerpts and supports filtering by website and publication date.

AI Gateway's web search capabilities have received a major update: Parallel's fast mode is now available for free, and Parallel has become the default web search tool for eve agents. For developers building retrieval-augmented applications and agent workflows, this means getting filterable, citable real-time web results without applying for a separate API key.
Free Scope and How It Works
According to the official announcement, Parallel's fast web search is free on AI Gateway and will remain so until January 31 of next year. This capability works with any model on AI Gateway, and developers don't need a separate Parallel API key—just declare it in a tool call.
It's important to note that the free offering covers only fast mode. Other search modes are still billed at standard rates, and the free mode enables Zero Data Retention by default—a practical configuration for teams sensitive to data compliance.
Zero Data Retention is a data handling policy: after processing a request, the provider does not persistently store the request content and returned results on its servers (typically retaining brief operational logs for billing and abuse detection). For enterprises, it reduces the risk of sensitive data being retained by third parties and triggering compliance issues, and it's often used as an entry requirement in GDPR, data localization, or internal security reviews. Note that zero data retention only constrains the search provider side—if your application writes search keywords or returned links to databases, logs, or analytics platforms, that's your own retention and requires separately designed retention periods and access controls per internal compliance requirements.
Parallel as eve's Default Search
Beyond the gateway-level availability, Parallel has also been set as the default web search for eve agents. In eve, the built-in web_search tool automatically uses Parallel's fast mode when the model is served through AI Gateway.
The real value here is "out of the box": previously, agents that needed web search often had to handle search provider accounts, key management, result parsing, and more. Now the default path is built in, and developers can focus on prompt design, post-processing, and business logic. The eve documentation also lists other search configuration options for swapping as needed.
Calling Parallel Search in AI SDK
Developers using AI SDK can mount the search tool via gateway.tools.parallelSearch. The official example combines the search tool with generateText and controls the number of multi-step reasoning rounds with stopWhen: stepCountIs(3):
import { gateway, generateText, stepCountIs } from 'ai';
const { text } = await generateText({
model: 'openai/gpt-5.6-terra',
prompt: 'What are the latest developments in AI this week?',
tools: {
parallel_search: gateway.tools.parallelSearch({ mode: 'fast' }),
},
stopWhen: stepCountIs(3),
});
There are two key points in this code: first, the search tool participates in model reasoning as an ordinary tool, and the model can decide on its own when to retrieve; second, stepCountIs limits the upper bound of multi-step calls, preventing the retrieval-generation loop from spiraling out of control and driving up cost and latency.
Tool calling refers to a mechanism where the model proactively outputs a structured call request during generation, the runtime executes that function, and the result is fed back to the model to continue reasoning. A complete "ask—retrieve—regenerate" cycle consumes multiple reasoning steps, and stopWhen with stepCountIs(n) sets a step limit: after n steps, the model stops calling tools and returns the current result. Without a limit, the model may repeatedly retrieve without converging, making cost and latency unpredictable. In practice, it's recommended to cap at 2–4 steps, explicitly instruct in the prompt to "answer directly when information is sufficient," and log each step's tool inputs and returns at the application layer to help diagnose retrieval quality issues.
Capability Boundaries and Retrieval Characteristics
Parallel's positioning is not "give an answer" but "find relevant web pages and return links with brief excerpts." This result format is better suited for scenarios that need source citations, fact-checking, or further processing.
It also supports filtering results by website or publication date. For time-sensitive content (such as industry news or product launches) and tasks that need to lock onto specific sources (such as searching only official documentation or designated media), filtering capability directly determines retrieval quality. Passing "brief excerpts + links" to the model, combined with filter conditions, is usually more controllable than letting the model guess sources on its own.
Retrieval-Augmented Generation (RAG) typically has two implementation paths: first, building a vector index in advance for similarity recall over private or relatively static corpora; second, calling web search at runtime for real-time results from the public internet. The former is highly controllable with stable latency but has lagging corpus updates; the latter has good timeliness and broad coverage, but result quality depends on the search service and query rewriting, and requires additional citation and verification logic. Parallel falls into the latter category, and it is not mutually exclusive with vector retrieval—a common approach is to first use web search to get the latest facts, then supplement with an internal index for proprietary knowledge, and finally require the model in the prompt to cite sources, balancing timeliness and accuracy.
What Developers Should Watch For
In the short term, this free policy lowers the cost of trial and error: you can mount the search tool directly in existing AI Gateway projects to validate an agent's retrieval-augmented performance without going through procurement first. In the medium term, Parallel becoming eve's default search means search is shifting from an "add-on capability" to part of the agent infrastructure.
Two things need advance planning. First, after the free window ends, fast mode's billing and the price gap with other modes will affect calling strategy—it's advisable to make the search mode a configurable option in your code. Second, data retention policy: the default zero data retention is compliance-friendly, but if your business needs to retain search records for auditing or performance analysis, you'll need to persist them at the application layer.
FAQ
Does the free offering cover all search modes? No. It's limited to fast mode; other modes are billed at standard rates.
Do I need a Parallel account to use it? Using it through AI Gateway requires no separate Parallel API key—the tool works directly with gateway models.
How do I switch to other search configurations in eve? The built-in web_search automatically uses Parallel's fast mode when the model is served through AI Gateway; refer to the eve documentation for other configuration options.
Can search results be filtered? Yes, by website or publication date, and returned content consists of links plus brief excerpts.
One-line Summary
Parallel's fast web search is free on AI Gateway and is now the default search tool for eve agents.
Paragraph Summary
AI Gateway rolls out a free policy for Parallel web search: fast mode is free from now until January 31 of next year, covering any model on the gateway. Developers don't need a separate Parallel API key—just declare it in a tool call—and zero data retention is enabled by default. Meanwhile, Parallel is set as the default web search for eve agents, and the built-in web_search tool automatically uses fast mode when the model is served through AI Gateway, eliminating account, key, and result-parsing setup. AI SDK users can mount search via gateway.tools.parallelSearch({ mode: 'fast' }) and limit multi-step calls with stepCountIs. Search returns links with brief excerpts and supports filtering by website and publication date.
Related articles

AI Writes Your Code—Do You Still Need to Learn Git? Five Reasons
AI coding agents handle Git operations automatically—is it still necessary to learn Git? Five reasons: team communication, reviewing agent PRs, monitoring code changes, fallback when AI goes down, and the joy of learning—plus how the focus of learning Git is shifting in the agent era.

Anthropic Reportedly Prices IPO at $83 Per Share, Valuing Company at $1.5 Trillion
Anthropic reportedly prices IPO at $83 per share, above the expected range, with a fully diluted valuation of $1.5 trillion. This article analyzes the signal significance of the pricing and its impact on the AI industry's financing environment.

Open-Source Code Hosting Platform Codeberg Suffers Outage, Sparking Debate in Developer Community
Open-source code hosting platform Codeberg.org suffered an outage, sparking Hacker News debate over nonprofit platform stability and open-source infrastructure dependency risks.