为所有模型提供一致的网络搜索与抓取能力

Postsby David Bai

Consistent Web Search and Fetch Across Every Model

Introducing openrouter:web_search

and openrouter:web_fetch

, two new tools that any model can call during a request. When a model decides to use one, OpenRouter executes it server-side and returns the result to the model without requiring any client-side implementation.

Web Search: a tool for agentic search 0 to N times per request, letting the model choose its own queries and timing.Web Fetch: a tool for retrieving full page content from any URL. Commonly used for pages found during search.

Try it now in the chatroom(opens in new tab) by clicking the tool icon

read the docs(opens in new tab) for the API details.

Swap Models Without Swapping Tools

Each model provider has its own built-in web search tool with a different schema. Switch models or providers, and you're stuck rewriting how you define, configure, and parse search results. You're also not guaranteed to have the same behaviors available, which can be problematic if you need features like strictly enforced blocked domains.

These new server tools give you one consistent way of enabling search and fetch. Specify {"type": "openrouter:web_search"}

once, and the tool definition, invocation, and result format stay identical across all tool-calling models. If you want identical search behavior as well, you can specify a provider like Exa or Parallel so the results coming back to the model are consistent regardless of whether the request routes to GPT-5.5, Claude, or Kimi.

Web Search

Web search supports four engines:

Each engine has different strengths. Native search is tightly integrated with the provider's model. Exa and Parallel add configurable result context size (search_context_size

), which native engines ignore. Most engines support domain filtering (allowed_domains

, excluded_domains

).

You can configure this in the chatroom UI or via the API:

Parallel Searches in Agentic Loops

When a model needs to compare information across sources, it can fire multiple searches in a single request. A question like "compare the pricing of the top 3 cloud GPU providers" might trigger three separate searches, each with different queries, before the model synthesizes an answer.

Use max_total_results

to cap cumulative results across all searches in a request. This keeps costs and context usage predictable:

Once the cap is hit, the model gets a message saying the limit was reached instead of running another search.

Web Fetch

Web fetch lets models retrieve full page content from URLs and comes with four supported engines.

Auto (default) | Uses native if supported, otherwise Exa | Varies |
Native | The provider's built-in fetch | Provider pricing |
OpenRouter | Direct HTTP fetch by OpenRouter | Free |
Exa | Content extraction and clean markdown output | $0.001 per fetch |

Specifying Exa or OpenRouter as the engine ensures consistent fetch behavior across all models, including the ability to restrict which URLs the model can fetch using allowed_domains

and blocked_domains

. Native provider fetch capabilities vary, so choose one of these engines if you need the parameters to be respected across models.

Use max_content_tokens

to cap how much content the model receives (useful for large pages that would eat your context window):

Migrating From the Web Search Plugin

Until now, models could only search through the web search plugin(opens in new tab), which ran exactly one search per request regardless of what the model actually needed. The model had no say in when to search, what to search for, or whether to search at all.

To migrate, replace plugins

with tools

in your request body:

Before (plugin):

After (server tool):

Server tools let the model decide when and how often to search. One caveat: server tools require a model that supports tool calling. If your current model doesn't support tools, you'll need to switch to one that does or keep using the plugin.

We've created a migration guide(opens in new tab) with full details.

链接抓取：https://openrouter.ai/announcements

Consistent Web Search and Fetch Across Every ModelGive any tool-calling model the ability to search the web and fetch page content on its own, with multiple search engines and fetch engines to choose from.May 7, 2026

GPT-5.5 Price Increase: What It Actually CostsOpenAI doubled per-token prices with GPT-5.5 but the model is less verbose. We measured real usage to see the net cost impact.May 4, 2026

New Audio APIs for Speech and TranscriptionText-to-speech and transcription are now live on OpenRouter. Two new endpoints give you access to speech synthesis and audio transcription across multiple providers, under one API.May 1, 2026

Response Caching: Zero Cost for Identical RequestsIntroducing the new Response Caching header: enables caching identical API requests so responses come back in a tiny fraction of the time, at zero cost.April 30, 2026

April Release SpotlightVideo generation, workspaces, an agent SDK, reranker models, and a wave of frontier model launches.April 30, 2026

Create OpenRouter Accounts via CLI with Stripe ProjectsRun stripe projects add openrouter/api to get an OpenRouter account, an API key, and Stripe billing, all from the command line. Your agents can do it too.April 29, 2026

Opus 4.7's New Tokenizer: What It Actually CostsAnthropic changed the tokenizer in Opus 4.7. We looked at usage that shifted from 4.6 to 4.7 to measure exactly how it affects costs.April 27, 2026

Build Your Own Harness with the Agent SDKUse the create-agent-tui and create-headless-agent skills to scaffold a personalized coding agent in minutes — with a terminal UI or headless for scripts and pipelines.April 24, 2026

Agent SDK: Building Multi-turn Agent Workflows on OpenRouterThe OpenRouter Agent SDK gives you callModel: one function that turns a chat completion into a multi-step agent with tool calls, stop conditions, and cost tracking across 300+ models.April 24, 2026

Introducing WorkspacesOrganize your OpenRouter projects into separate environments, each with its own API keys, routing defaults, guardrails, and observability.April 22, 2026