Serve every model with frontier-level knowledge

Inference providers call Exa Search server-side and offer web search as a built-in feature, so developers have access to native search without needing a second vendor key. Exa returns cited passages at answer time, which lets open models answer from today's web. OpenRouter uses Exa as the default search for models without built-in search.

LLM grounding gives a language model external sources to answer from, so its responses are evidenced from the latest information instead of from training data alone. Exa grounds LLMs in the live web: at answer time the model searches Exa and composes its response from the cited information Exa provides.

A knowledge cutoff is the date a model's training data ends. Anything published after that date is missing from the model. Search fills that gap. The model calls Exa when a question needs newer information and answers from today's pages.

Exa is OpenRouter's default web search for models without built-in search, such as open-weight models from DeepSeek, Meta's Llama, Qwen, and Mistral. When a developer adds :online to a model or calls the openrouter:web_search tool, Exa runs the search and returns highlights that OpenRouter passes to the model.

Set the search type on each request. The instant type returns results in under 250 ms, fast in about 450 ms, and auto, the default, in about 1 second. For research questions, deep-lite (about 4 seconds), deep (4 to 15 seconds), and deep-reasoning (12 to 40 seconds) search in several steps and combine the results.

Yes, without retraining the model. In Exa's published RL study of May 13, 2026, holding the model fixed and changing only the search backend raised SimpleQA Pass@1 from 0.579 to 0.651 on an untrained Qwen3-4B and from 0.689 to 0.730 on an untrained Qwen3-235B. Exa scored higher on 11 of the 12 model and benchmark pairs the study reports. Setup and full tables are in the RL search outcomes post.

Not on their own, but a host can add search by giving the model Exa as a tool. OpenRouter uses Exa as the default web search for models without built-in search, and Baseten grounds open-source models with Exa through Baseten Hosted Tools.

Per request or by contract. Exa Search starts at $4 per 1,000 requests on pay-as-you-go, and Enterprise agreements offer volume pricing, custom rate limits, ZDR, and SLAs. Providers that wish to offer Exa natively can learn more with our partner program.

Yes. Exa is SOC 2 Type II audited and honors GDPR and CCPA data-subject rights. Enterprise plans offer zero data retention, HIPAA with a BAA, SSO and SCIM, and a custom MSA and DPA. To access audit reports, request them in the Trust Center or talk to our team.