Grok by xAI:
Complete Guide to Models,
Features and Pricing

Grok is a conversational AI assistant developed by xAI. It lives in three places: the standalone web and mobile app at grok.com, inside X (formerly Twitter) for X Premium subscribers and above, and through a developer API at api.x.ai. The current flagship version is Grok 4.3, released April 30, 2026, with a 1M token context window and native video input. Older variants including Grok 4 (256K), Grok 4 Fast (2M), Grok 4.1, Grok 4.20, and Grok 3 remain accessible through the API.

Listen to this research in a podcast mode

The name comes from Robert Heinlein’s 1961 novel Stranger in a Strange Land, where “to grok” means to understand something deeply and intuitively. The name is shared with an open-source log-parsing library and used as a verb, but for purposes of this guide and search disambiguation, “Grok” refers specifically to xAI’s assistant.

What distinguishes Grok from other frontier AI assistants is access pattern, not architecture. Grok is the only major model with a native real-time stream from X, and the only consumer-accessible model with a 2M token context window on its Fast variants. It also accumulates the most public controversy of any frontier model in this generation, including a July 2025 incident where it produced antisemitic content at scale. Both characteristics are documented and both shape practical use.

xAI is an AI company founded by Elon Musk in July 2023. The company’s stated mission is “to understand the true nature of the universe.” It is headquartered in Palo Alto, California, with primary training infrastructure at the Colossus data center cluster in Memphis, Tennessee.

In March 2025, xAI completed an all-stock acquisition of X (formerly Twitter), valuing xAI at $80 billion and X at $33 billion. The merger gave Grok structural access to X’s content stream. A separate report from February 2026 referenced an xAI-SpaceX merger via an X post attributed to @Grok; corporate structure details require primary verification and are not yet documented in xAI filings.

xAI’s reported valuation was approximately $200-230 billion as of January 2026, following a Series E round of around $20 billion fueled by Middle Eastern sovereign capital. Total funding raised across rounds is reported at approximately $45 billion. Co-founder Igor Babuschkin (formerly DeepMind) handles much of the technical communication. Linda Yaccarino departed as X CEO in summer 2025.

Colossus operates at approximately 1-2 GW with 200,000 to 555,000 NVIDIA GPUs across two facility expansions, depending on the disclosure date. xAI has been more transparent than most frontier labs about training infrastructure, less transparent about model architecture details such as parameter counts and expert configurations.

xAI’s stated design principle for Grok is “truth-seeking.” In practice, this resolves into three product behaviors that you can observe across versions: a willingness to engage controversial topics other models refuse, a conversational personality that leans toward direct and irreverent rather than cautious, and a system prompt history that has explicitly instructed the model to make politically incorrect claims when “well substantiated.” That last instruction was removed from the public xAI GitHub system prompts after the July 2025 antisemitic content incident.

What this means for users is a model that attempts more answers than peers refuse. Across independent benchmarks, this shows up as a high “answer rate” combined with a high error rate when the model is uncertain. On the AA-Omniscience benchmark, Grok 4 attempts answers it should refuse 64% of the time. Claude 4.1 Opus, for contrast, achieves a 0% rate on the same metric by declining when uncertain. Both are valid design choices. They produce different failure modes.

In multi-model evaluation, Grok’s behavior matches its design intent. Per the Suprmind Multi-Model Divergence Index, April 2026 Edition (n=1,324 production turns), Grok surfaces 509 unique insights (19.7% share, third among five providers) that the consensus models miss. The trade-off is that its calibration delta on high-stakes turns is only -1.9 points: it does not measurably hedge when the question carries more weight. The contrarian insights arrive with the same apparent confidence as the incorrect ones.

xAI has released six generations of Grok models since November 2023. The current active lineup centers on the Grok 4 family (Grok 4, Grok 4 Fast, Grok 4.1, Grok 4.20, Grok 4.3) plus older Grok 3 and Grok 2 variants in the API. The flagship recommendation in xAI’s official docs is Grok 4.3.

Active Grok Models in 2026

The variant matrix below covers every model currently accessible through grok.com or the API. Context windows refer to input tokens. API IDs are the strings developers pass to the Chat Completions endpoint.

Sources: xAI official docs (docs.x.ai/docs/models, accessed 2026-04-16); per the Suprmind Multi-Model Divergence Index, April 2026 Edition; per Suprmind’s AI Hallucination Rates and Benchmarks reference (May 2026 update).

Grok 4 vs Grok 3: What Changed

Grok 3 introduced DeepSearch, DeeperSearch, Think mode, and reinforcement learning at post-training. Grok 4 moved RL into pretraining scale (10x compute over the previous RL run), introduced multi-agent Heavy configurations, native voice, and camera mode, and pushed context to 256K. Grok 4 Fast extended that to 2M tokens at $0.20/$0.50 per million tokens, the first xAI model to reach the 2M threshold and the lowest API price point in the family.

The benchmark trajectory is mixed. On Vectara summarization hallucination, Grok 3 scored 2.1% (excellent) on the old dataset. Grok 4 scored 4.8% on the same dataset and over 10% on the harder new dataset. On Columbia Journalism Review citation accuracy, Grok 3 scored 94% citation hallucination, the worst of any model tested in that study. Grok 4 has not been independently retested on CJR at the time of this guide.

Grok 4.20 Reasoning: The Calibration Story

Grok 4.20 Reasoning is the variant in the family with the calibration improvement story. On the Artificial Analysis AA-Omniscience benchmark, it scores 17% on the “when attempting” hallucination rate – the lowest rate among Grok variants tested at that time, and a meaningful drop from Grok 4’s 64% and Grok 4.1 Fast’s 72%. Per Suprmind’s AI Hallucination Rates and Benchmarks reference, this is the first Grok variant to demonstrate measurable calibration improvement.

For workflows where a wrong answer costs more than no answer, Grok 4.20 Reasoning is the variant to specify. It is available in the API as grok-4.20-reasoning at $2/$6 per million input/output tokens (Artificial Analysis) – a separate independent source (TheRouter) reports $3/$9, with the conflict unresolved at publication.

What Is Grok 5?

Grok 5 has been referenced repeatedly by Elon Musk and xAI’s official X account as the next major architectural step. Per Fello AI citing xAI’s X account (May 2026), Grok 5 is targeted for Q2 2026 public beta after the Q1 2026 target slipped. MindStudio (April 30, 2026) reports xAI is training parallel Grok 5 variants ranging from 6 trillion to 10 trillion parameters per Musk’s public statements; primary source is not directly linked. Grok 4.4 (~1T parameters) is reported 2-3 weeks from late April 2026; Grok 4.5 (~1.5T) is reported 4-5 weeks out. Treat all Grok 5 timing as Volatile – verify at xAI’s official X account before publication or planning.

Grok has six consumer tiers, two business tiers, and a tiered API. The structure rewards close reading because tier names do not map cleanly to model versions, and tier-to-model assignment changes during staged rollouts. The honest pricing question for most users is not “how much does Grok cost” but “which Grok model do I actually get on which tier.”

Consumer Tiers

X Premium ($8/mo) is omitted from the highlights above; full tier details for all six consumer tiers are documented in the pricing guide. Sources: felloai.com (May 2026); fritz.ai (January 2026); TechCrunch (July 2025, SuperGrok Heavy launch).

SuperGrok vs X Premium+: When Each Makes Sense

SuperGrok at $30/month is a Grok-focused subscription. X Premium+ at $40/month bundles Grok with X platform features (reduced ads, longer posts, monetization). Same model access, different value bundle. Pick SuperGrok if Grok is the primary use case. Pick X Premium+ if you would buy X Premium+ anyway.

SuperGrok Heavy: Who It Is For

SuperGrok Heavy at $300/month is the only consumer tier with confirmed full Grok 4.3 access (lower tiers receive Grok 4.3 in staged rollout). It also opens access to the 16-agent parallel mode used in Grok 4 Heavy benchmark demonstrations. The $300 ceiling restricts the tier to professional and enterprise users by cost alone.

Grok API Pricing

Pricing conflict notes: Grok-4.20-reasoning is reported at $2/$6 by Artificial Analysis and $3/$9 by TheRouter. We use Artificial Analysis as the authoritative independent source. Verify at console.x.ai before publication. Grok-4.1 pricing is not displayed on the docs.x.ai pricing page as accessed in research; rates are from third-party aggregators.

API tools are billed separately: web search, X search, code execution at $5 per 1,000 calls each; file attachments at $10 per 1,000; Collections search at $2.50 per 1,000. xAI offers up to $175/month in free API credits for new accounts.

What Model Do You Actually Get on Each Tier?

This is the documented opacity. SuperGrok at $30/month is described as “Grok 4.3 rolling out in stages.” Tier-equivalent users receive different models simultaneously, with no UI indicator of which model processed any given query. Auto Mode compounds this by routing dynamically across model variants without disclosure. The only firm disambiguation path is the API, where developers can pin specific dated model IDs (e.g., grok-4-0709).

For SuperGrok Heavy users at $300/month, full Grok 4.3 access is confirmed. For SuperGrok and X Premium+ users at $30-40/month, the model assignment is partially staged. For Free and X Premium users at $0-8/month, the model is Grok 4 with reduced context and rate limits, sometimes routed to older variants. None of this is exposed in the consumer UI as of publication. If your workflow depends on knowing which model answered, use the API with a dated model ID.

For deeper coverage of tier-to-model mapping, see the Grok Pricing Guide →

Grok ships with a feature set that overlaps with other frontier assistants on the basics (chat, voice, image generation) and diverges on a few items unique to xAI (real-time X access, Companions, the multi-agent Heavy configuration). The features below are organized by use case.

For parser fidelity notes, OCR behavior, and full feature mechanics, see the Grok Features Deep Dive →

Grok’s benchmark profile is the most divergent of any frontier model family. xAI publishes results that position Grok at or near the frontier; independent evaluation platforms show materially different numbers depending on the failure mode being measured. This is not a contradiction. Different benchmarks measure different things, and Grok’s performance varies enormously across them.

How to Read Grok’s Benchmark Profile

Grok’s reliability profile splits cleanly into four measurement categories. Each one tests a different failure mode. A model can score excellent on one and poor on another, and both numbers are accurate.

• Vectara HHEM measures summarization faithfulness. Does the model add facts not in the source document?
• AA-Omniscience measures knowledge calibration. When the model does not know something, does it admit uncertainty or fabricate?
• FACTS measures multi-dimensional factuality including search-grounded and multimodal accuracy.
• Columbia Journalism Review (CJR) measures citation accuracy. Are cited claims actually in the cited sources?

Grok-3 scored 2.1% on Vectara (excellent) and 94% on CJR (worst of any model tested). Same model. Same era. Both numbers accurate. They tell different parts of the same story.

Hallucination Rates Across Grok Variants

Sources: Vectara HHEM Leaderboard (2026); Artificial Analysis AA-Omniscience (Feb 2026); Google DeepMind FACTS (Dec 2025); Columbia Journalism Review (Mar 2025).

For full cross-model comparison and methodology, see Suprmind’s AI Hallucination Rates and Benchmarks reference →

Grok on Citation Accuracy (CJR)

Grok-3 scored 94% citation hallucination on the Columbia Journalism Review citation accuracy test. The worst score of any model tested. By comparison, Perplexity Sonar Pro scored 37%, ChatGPT scored 67%, Gemini scored 76%. This is not a caveat at the bottom of a review. It is a structural constraint that defines where Grok can and cannot be deployed alone.

The conditions that trigger citation hallucination are not unusual: any task requiring source attribution including research synthesis, journalism support, literature review, and citation-grounded analysis. Grok does not need to be doing something exotic for the failure to appear. For citation-dependent work, pair Grok with a model that has stronger attribution discipline – Perplexity is the cleanest pair on the data.

The Internal vs Independent Benchmark Divergence

The Grok 4.1 Fast story is the most flagged. xAI claimed a 65% hallucination reduction from Grok 4 to Grok 4.1 Fast on internal benchmarks (12.09% to 4.22%). AA-Omniscience independently measured Grok 4.1 Fast at 72% – worse than Grok 4’s 64%. The MASK sycophancy benchmark also increased (0.07 to 0.19-0.23). Both data sources are accurate. They measure different things.

The Grok 4.20 Reasoning calibration improvement is the most underreported finding. At 17% on AA-Omniscience’s “when attempting” metric, it is the first Grok variant to show meaningful calibration improvement. For workflows where a wrong answer costs more than no answer, this is the Grok variant to specify.

The takeaway is not that xAI’s benchmarks are wrong. They measure what they say they measure. The takeaway is that the configuration matters: a Heavy multi-agent score is not directly comparable to a single-model score from a peer vendor, and a benchmark tuned for a specific evaluation harness is not the same as performance in a production workflow.

The comparison stories are different for each peer. Against ChatGPT, Grok wins on speed and real-time data and trails on enterprise maturity. Against Claude, Grok wins on context window size and trails on calibration. Against Gemini, the two models disagree more than any other pair in the multi-model dataset. Against Perplexity, Grok has a real-time X stream but trails on citation accuracy.

Five-Model Snapshot

Per the Suprmind Multi-Model Divergence Index, April 2026 Edition (n=1,324 production turns).

For deeper head-to-head with structured benchmark comparison and use-case decision tables, see Grok vs Other AI Models →

Grok accumulates the most documented public controversy of any frontier AI model in this generation. Three controversies are the most widely reported, and three regulatory actions are active. The facts below are current to the May 2026 research pass.

Five orchestration patterns where
Grok adds signal an ensemble needs.

Grok’s value is highest when it is one model in an ensemble, not when it is treated as a sole-model oracle. The five orchestration patterns below come from documented data on where Grok adds signal and where it needs another model’s discipline as a counterweight.

For full detail on Grok’s behavior across all five providers, see the Suprmind Multi-Model Divergence Index →