xAI (Grok) product ecosystem — Library index

Canonical reference for everything xAI ships under the Grok brand: the model lineup (Grok-1 open-weight release → Grok-1.5/1.5V → Grok-2 → Grok-3 with Think/DeepSearch → Grok-4 / Grok-4 Heavy → Grok 4.1 → and the fast-moving 4.x point releases beyond the internal cutoff), the xAI API (an OpenAI-compatible /v1/chat/completions surface plus the newer /v1/responses endpoint, function calling, structured outputs, vision, and Live Search real-time grounding), pricing / context / rate-limit tiers, distribution (api.x.ai, grok.com, native X/Twitter integration, Azure AI Foundry hosting), and the Colossus Memphis training-cluster story that underwrites the release cadence. House style and source-citation discipline match the Claude/, OpenAI/, and GoogleGemini/ sub-libraries; every non-obvious claim cites docs.x.ai, an x.ai/news post, an xai-org/* GitHub repo, or a named third-party source, and every fast-moving claim is date-stamped.

This sub-library lives inside Math-and-Compute/Compute/ alongside Claude/, OpenAI/, GoogleGemini/, and AWSBedrock/ — the other frontier-lab vendor sub-libraries — plus the broader compute-systems notes. xAI moves on the same monthly cadence as the rest of the frontier: as of the 2026-05-30 authoring pass, Grok-4 (mid-2025) and Grok 4.1 (Nov 17 2025) are the well-attested current generation, and a string of later point releases (Grok 4.2 onward, Grok 4.20, Grok 4.3, Grok-Build, the Grok Imagine media stack) surface in live xAI docs but land past the internal knowledge cutoff (~Jan 2026) — those are noted lightly here and flagged “verify” so a future reader re-checks them against docs.x.ai before acting.

The contents are a single flat folder of notes (matching the Claude sub-library rationale): most lookups are “what’s the context window / price of model X” or “how does Live Search work”, and a flat folder collapses to one Obsidian search.

Positioning — what makes xAI/Grok distinct

Three things differentiate Grok from the GPT / Claude / Gemini frontier:

1. Real-time X (Twitter) grounding. xAI owns X, so Grok has first-party, real-time access to the X firehose plus web search via the Live Search / web_search tool. This is the structural moat: live social signal and breaking-news recency that the other labs source through partnerships. DeepSearch (introduced with Grok-3, Feb 2025) packages this into an agentic research loop over the web and X.
2. “Maximally truthful / maximally curious” framing. xAI markets Grok as less filtered and more willing to engage with contested topics than its peers — a deliberate positioning contrast with the more guard-railed competitors. Treat this as marketing posture, not a measured safety claim; Grok 4.1 (Nov 2025) headlined a ~3× hallucination-rate reduction (12.09% → 4.22% on xAI’s own web eval), which is the substantive version of the “truthful” pitch.
3. The Colossus compute story. xAI’s Memphis Colossus cluster — built 0→100k Nvidia GPUs in 122 days (Sep 2024), doubled to 200k in another ~92 days, and reported at 150k H100 + 50k H200 + 30k GB200 by Dec 2025 — is the world’s largest single-coherent AI training cluster. The unusual cadence (Grok-2 → Grok-3 was “10× more compute”; Grok-4 leaned on RL-at-pretraining-scale on Colossus) is downstream of having that much homogeneous compute in one building.

See also

• _index — sister sub-library for Anthropic; the closest peer for cross-vendor comparison
• _index — sister sub-library for OpenAI (Grok’s API is OpenAI-compatible, so this is the migration reference)
• _index — sister sub-library for Google Gemini
• _index — multi-vendor hosting sub-library (Grok is hosted on Azure AI Foundry rather than Bedrock — see the availability note)
• llm-landscape — the vendor-neutral landscape catalog; the ### xAI row there is the at-a-glance version of this sub-library
• transformer-architecture — the MoE-transformer substrate Grok is built on (Grok-1 was a 314B MoE)
• fine-tuning-rlhf — RL / RLVR, the training axis behind Grok-3/4’s reasoning (“Think”) mode
• prompt-engineering-agent-systems — vendor-neutral agentic discipline overview
• cuda-triton-gpu-programming — the GPU-programming substrate for the H100 / H200 / GB200 silicon inside Colossus

Library contents

Foundations

This sub-library starts as a single deep note. As the surface grows (a dedicated Live-Search note, an Agent-Tools-API note, a Grok-Imagine media note), split them out as siblings under xAI/ and add rows here — mirroring how Claude/ and OpenAI/ fanned out.

How to use this library

1. Picking Grok for a new project — read grok-models-and-api §lineup for the model matrix and §“when to pick Grok” for the decision, then cross-check the live numbers on docs.x.ai/developers/models (pricing and the 4.x lineup move fast).
2. Migrating from OpenAI — the xAI API is OpenAI-compatible (base_url="https://api.x.ai/v1" with the openai client). grok-models-and-api §API has the drop-in pattern and the points where xAI diverges (Live Search tool, the /v1/responses endpoint, model IDs).
3. Real-time / news-grounded apps — this is Grok’s structural advantage; grok-models-and-api §Live-Search covers the web_search tool, X-firehose access, and DeepSearch.
4. Cross-vendor comparison — pair grok-models-and-api §benchmarks with claude-models-and-capabilities and openai-models-and-capabilities, and read the landscape framing in llm-landscape.

Conventions used in this library

• Model IDs are quoted verbatim (e.g. grok-4, grok-4-fast-reasoning, grok-3-mini, grok-4-1-fast). xAI uses dateless aliases that resolve to the latest stable snapshot; legacy aliases (grok-4, grok-3, grok-3-mini) increasingly route forward to the current flagship — pin and verify before relying on an alias.
• Pricing is in USD per million tokens (MTok). Where multiple \$ symbols appear on the same line they are backslash-escaped to avoid Obsidian MathJax interpretation.
• Equations and rate-limit math use fenced code blocks, never $-delimited math (repo KaTeX hard-fail convention).
• Code examples use the OpenAI Python SDK pointed at api.x.ai (the compatibility path most teams use) and the official xai-sdk-python where it differs.
• “Per docs.x.ai 2026-05” tags a claim sourced from official documentation as of this file’s created date. “Per x.ai/news/grok-4 (Jul 2025)” tags an xAI announcement. Third-party numbers are attributed to the named source (VentureBeat, NVIDIA newsroom, Microsoft Azure blog, Wikipedia, SemiAnalysis).
• Releases dated after ~early 2026 (Grok 4.2 / 4.20 / 4.3, Grok-Build, Grok Imagine) are past the internal knowledge cutoff (~Jan 2026) and are labelled “verify against docs.x.ai” so a future reader re-checks them.

Sources read while building this library (2026-05-30 pass)

• docs.x.ai/developers/models (live model catalog + pricing + context windows)
• docs.x.ai/overview (API base URL, auth, /v1/responses endpoint, OpenAI compatibility)
• docs.x.ai/docs/guides/live-search (Live Search / web_search tool)
• docs.x.ai/docs/guides/function-calling + docs.x.ai/docs/guides/structured-outputs (tool use + JSON-schema mode)
• github.com/xai-org/grok-1 (Grok-1 open-weight release — 314B MoE, Apache-2.0, Mar 2024)
• github.com/xai-org/xai-sdk-python (official Python SDK)
• x.ai/news/grok-4 (Grok-4 launch + benchmarks, Jul 2025) and x.ai/news/grok-4-1 (Grok 4.1, Nov 2025)
• x.ai/colossus + Wikipedia “Colossus (supercomputer)” + NVIDIA newsroom Spectrum-X release + SemiAnalysis “Colossus 2” (infrastructure)
• Wikipedia “Grok (chatbot)” (lineage timeline + dates)
• VentureBeat (Grok 4.1 hallucination-rate launch coverage) + Microsoft Azure AI Foundry blog (Grok hosting)

Fast-moving claims are date-stamped inline; anything past the ~Jan 2026 internal cutoff is flagged “verify” in grok-models-and-api so it can be re-checked against the live xAI docs.