How Grok Works:
DeepSearch, Think Mode,
Companions and More

DeepSearch is the feature that turns Grok from a chat model into a research agent. Activated through a UI toggle on grok.com or by prefixing prompts with “Use DeepSearch:”, it fires an iterative retrieval-augmented-generation loop. The agent splits the query into sub-queries, runs parallel searches against the web and X, follows fresh links, summarizes each batch in an internal scratchpad, and repeats until it hits a 10-step limit or a time threshold.

The model cross-checks up to seven consistency layers before drafting the response. Users can toggle a “Thoughts” view to see the intermediate reasoning steps. In the API, DeepSearch maps to enabling the web_search and x_search server-side tools. Citations are only generated when these tools are invoked.

DeeperSearch is the more thorough variant. It runs additional iterations, traverses deeper through linked sources, and produces a longer synthesis stage. xAI employees described it as “an improved version of DeepSearch” that goes “two steps further.” Latency is the trade-off: DeeperSearch takes meaningfully longer.

Think Mode activates Grok’s reasoning model path. Instead of producing an answer in one shot, the model generates an internal chain-of-thought, visible via the “Thoughts” toggle, before producing output. xAI’s official description: Think “focuses on advanced reasoning and problem-solving… like a human thinking.” Available across tiers for basic Think access; Heavy mode (extended reasoning) requires SuperGrok Heavy. Grok 4.3 has reasoning always on by default with no toggle.

Why reasoning increases summarization hallucination

The mechanism is documented. Reasoning models invest compute into generating inferences. When the task is open-ended analysis, those inferences add value. When the task is grounded summarization (compressing a source document into a shorter version), the same inference engine adds conclusions and inferences the source document does not contain. Vectara measures this as fabrication – the model added facts not in the source.

The practical guidance: turn Think Mode on for analytical tasks where step-by-step logic matters and you want chain-of-thought transparency. Turn it off for document summarization, citation-grounded research, and any task where the deliverable is compressed source material. The reasoning tax is real and quantified.

When Think Mode earns its cost

Open-ended technical analysis. Multi-step problem decomposition. Math problems where intermediate steps validate the conclusion. Strategic decisions where you want to inspect the model’s reasoning before accepting the answer. In these cases, the visible chain-of-thought is signal, and the inference-heavy generation produces value rather than fabrication.

Expert Mode is a usage mode rather than a tier or a model version. Users select it manually from the consumer app. It forces higher compute and deeper reasoning regardless of query complexity, in contrast to Auto Mode which routes dynamically.

The gap: no verbatim official xAI definition of Expert Mode appears in xAI’s published documentation as of this guide’s research pass. We searched docs.x.ai, the xAI blog, the Grok user guide, and primary launch announcements. Expert Mode appears in third-party YouTube tutorials, third-party review articles, and the consumer app UI itself, but not in xAI’s own documentation as a defined feature with stated mechanics.

Third-party descriptions place Expert Mode in this hierarchy

• Auto Mode – dynamic routing based on perceived query complexity
• Fast Mode – quick response, no thinking step
• Expert Mode – higher compute, deeper reasoning
• Thinking Mode (Beta) – full RL reasoning with visible chain-of-thought
• Heavy Mode – 16-agent parallel architecture (SuperGrok Heavy only)

In this framing, Expert Mode sits between Fast (quick, no thinking) and Thinking (full reasoning), forcing a deeper compute path than Auto Mode would select for the same query.

The honest answer: Expert Mode produces more thorough responses than Auto Mode for the same query, with longer latency. It is available across consumer tiers including the free Grok 4 access window when xAI opened that in August 2025. The mechanics beyond that are not documented by xAI itself. If your workflow depends on Expert Mode behavior, treat its current implementation as Volatile.

Grok handles document upload and analysis through both the consumer app and the API. The supported format set covers most everyday workflows.

API model restrictions and parser fidelity

API document processing is restricted to Grok 4 and newer models. Grok 3 and Grok 2 do not support file uploads through the API. The chat UI handles document upload across all consumer tiers, with Free tier subject to per-window rate limits.

PDF parsing is confirmed for both surfaces. The model can execute Python code on uploaded files via the code execution tool when enabled. For structured data uploads (CSV, XLSX), keep row counts under 200,000 to avoid timeouts.

What is not formally documented: DOCX table extraction fidelity, embedded image extraction, footnote handling, and OCR behavior on scanned PDFs. If your workflow depends on these specifics, test empirically rather than relying on documentation that has not been published.

Collections and RAG

Documents can be stored in Collections – a persistent vector store – and queried via the collections_search tool at $2.50 per 1,000 calls. File storage is billed at $0.025/GiB/day; collection storage at $0.10/GiB/day. File storage charges began April 20, 2026. This makes Grok competitive on the document-grounded use case, with the caveat that the reasoning variant’s Vectara New Dataset hallucination rate (20.2%) means Think Mode amplifies summarization fabrication.

Imagine is xAI’s image and video generation surface, accessible separately from the chat API at a dedicated Imagine endpoint. Image generation through the Aurora model has been available since before Grok 4. Video generation rolled out with the Grok 4 launch period in July 2025.

Voice mode predates Grok 4 and was significantly upgraded with the Grok 4 launch on July 9, 2025. Camera mode launched simultaneously: users point their camera and Grok analyzes the visual scene while speaking.

The voice model is trained in-house using xAI’s RL framework and speech compression techniques. Pre-Grok 4 voice was described in independent reporting as “bolted-on rather than native”; the Grok 4 launch addressed that critique with a redesigned voice path. Basic voice on Free tier and above. Priority voice and camera mode require SuperGrok and above.

API voice pricing

The Text-to-Speech API launched in general availability in early May 2026, expanding the developer voice surface beyond Realtime voice and STT endpoints.

Companions are 3D animated AI characters launched alongside Grok 4 on July 14, 2025. The current character lineup includes Ani (anime-style female companion), Rudy (a friendly red panda), Bad Rudy (vulgar variant), and Valentine (male companion, teased by Musk on July 17, 2025). Users access Companions through a dedicated tab in the consumer apps.

Companions use the Grok 4 underlying model with real-time image generation and persistent memory across conversations. Each character has a distinct personality and conversation style. Some Companions support NSFW mode; Ani specifically has been documented appearing in lingerie on command.

For users who want a chat AI with a structured personality and conversation continuity, Companions are a documented feature. For users in professional contexts where the explicit-capable framing is a brand or compliance risk, the standard chat interface (with Memory and Projects) provides equivalent conversation continuity without the Companions framing.

Grok Memory is a documented consumer app feature launched in beta on April 19, 2025. Memory is stored outside the context window and persists across sessions; users can review, edit, and delete memory entries through settings. Memory is selectively injected at conversation start – only relevant context is retrieved, not loaded wholesale.

The mechanism documented in independent sources: when a user starts a new conversation, Grok queries its memory store for relevant prior facts and preferences, then injects those into the conversation context before generating a response. This reduces context-window pressure for long-running users while maintaining personalization.

User-facing limits

“Asking Grok to forget certain information does not automatically erase it” – manual deletion through settings is required. Memory often breaks across conversations (documented in independent reviews of Grok memory behavior). Memory was not available in certain regions at the April 2025 beta launch; specific regions were not disclosed.

For users who depend on memory continuity, the feature works in the consumer apps. For developers building memory-dependent applications, plan to build the memory layer yourself.

Projects (also called Workspaces) act as containers for related chats, files, and custom instructions. The official description from grok.com/project: “Supercharge Grok with Projects. Create custom workspaces, upload files for smarter chats, and collaborate securely.” Each workspace holds persistent files, conversation history, and custom prompts. Users train the workspace by uploading documents that subsequent chats can reference. Workspaces launched approximately April 12-15, 2025.

Grok Build is a coding agent in pre-launch as of May 2026. Reports from January 2026 described “early look” coverage. xAI announced Build and a companion CLI tool launching in mid-to-late April 2026 (precise date not confirmed). Full public launch was not confirmed at the research date.

Build is a separate feature from Imagine and is not part of the standard chat interface. It is positioned to compete with Claude Code and similar coding agents from other vendors.

Grok’s citation system is feature-conditional. Citations are generated when server-side search tools (web_search, x_search, attachment_search, collections_search) are invoked. Without these tools enabled, Grok produces no citations even when responding to factual queries that would benefit from source attribution.

When tools are enabled, the agent records all accessed URLs and attaches citation metadata to relevant portions of the answer. In the UI, citations appear as inline clickable links. In the API, citations are returned as structured fields in the response. The return_citations: true parameter ensures URL list return.

The source mixing problem

Sources are mixed: web URLs and X posts are labeled, but no systematic distinction is visually enforced between X-sourced and web-sourced claims in most UI presentations. This matters because X content quality varies enormously. A peer-reviewed paper and a viral X thread can appear in the same citation list with similar visual treatment.

No documented per-query citation count limit exists. The 10-step DeepSearch limit caps the number of sources consulted per multi-step research task.

Tier availability for some features is not enumerated in official xAI docs as of May 2026. Treat tier-specific limits as Volatile and verify at grok.com/plans before relying on the cap for production planning.