Grok (xAI) for Automation

Key takeaways

• Cheap per token, strict about prompt size. Grok 4.7 (21 September 2026) costs $2 input / $0.50 cached / $6 output per million tokens and has a 500K context. On every current Grok model, a prompt of 200K tokens or more doubles the price of the whole request. The cheaper grok-4.3 and grok-4.20 cost $1.25 / $2.50 with a 1M context. Grok 5 has not shipped.
• A complete automation API. Grok offers function calling, schema-enforced structured outputs, server-side web search and X (Twitter) search, a Python sandbox, document collections (RAG), remote MCP tools, a batch API and realtime voice. All of it runs through an OpenAI-SDK-compatible Responses API. The Anthropic-compatible endpoint is deprecated. Batch discounts (20%) apply only to the cheaper grok-4.3 and grok-4.20 models, and the only regional option is a US endpoint for grok-4.7 and grok-4.6 at 1.1x.
• Capable, not a leader. Artificial Analysis scores Grok 4.7 at 46 against a median of 26. On AA's live leaderboard (v4.3.2, 12 October 2026), Claude Opus 5.5 (Max) scores 57.6, Claude Sonnet 5.5 56.0, GPT-6 Astra 52.7 and GPT-6.1 Sol 51.8, against 46.4 for Grok 4.7 (XHigh). Grok 4.7 is verbose, so each benchmark task costs 2.5 times as much as on Grok 4.6 at the same token price.
• A big platform that loses money. xAI is now SpaceX's AI division (SpaceXAI), and SpaceX has been listed on Nasdaq since 12 June 2026. The AI segment's 2025 revenue was $3.2 billion, against a $6.4 billion operating loss. Grok's AI features had 117 million monthly users at 31 March 2026.
• Deep US government ties. These include a $200M-ceiling Pentagon contract (July 2025), IL5 accreditation on GenAI.mil (August 2026) and a GSA OneGov deal offering Grok 4 and Grok 4 Fast at $0.42 per agency for 18 months, valid until March 2027.
• The heaviest regulatory baggage of any frontier model. After the January 2026 sexual-deepfake scandal, Indonesia, Malaysia and the Philippines blocked Grok for 5 to 22 days. India's IT ministry (MeitY) issued notices, the EU and the UK opened investigations, and Canada found a privacy-law breach. Most of these cases are still open.
• Verdict. Use Grok for agents that need live X data, low-cost voice, structured extraction under 200K tokens, and as a second vendor. Avoid it for images of real people, for workloads that need India or EU data residency, for zero-retention RAG, and as the only model in brand-sensitive consumer products.

All facts are as of 12 October 2026; prices, benchmarks, cloud availability, filings and court dockets were re-checked against primary sources on 12 October 2026. They are drawn from vendor documentation, SEC filings, court dockets, Wikipedia and news coverage. Labels:

• Vendor-reported: published by xAI or another vendor about itself.
• Wikipedia-only: the only source read was a Wikipedia article, whose references were not opened.
• Secondary: news or blog coverage not checked against a primary source.
• Single-source: one source only.

Prices are in US dollars per million tokens. For the wider field, see Top 10 models compared and the AI automation landscape.

Six new models since March 2026, and a forced migration in May

SpaceX's acquisition of xAI closed on 2 February 2026, all stock, at a fixed 0.1433 SpaceX shares per xAI share (before SpaceX's 2026 split), with a $75.46-per-share cash election for eligible service providers (SpaceX 424B4; Q2 2026 10-Q). In May 2026 Musk said xAI would cease to exist as a separate company. In July it was rebranded SpaceXAI (Wikipedia: SpaceXAI); the name "SpaceXAI LLC" appears in an 8th Circuit docket caption (CourtListener). The developer docs now show "SpaceXSI" in their page title, and the two spellings have not been reconciled (xAI models page). This document says "xAI" for the developer and "Grok" for its models.

The 2025 models were all retired at 12:00 PM PT on 15 May 2026, together with grok-imagine-image-pro:

• Grok 3 (17 February 2025; API in April 2025 at $3 / $15)
• Grok 4 and Grok 4 Heavy (9 July 2025)
• Grok Code Fast 1 (28 August 2025)
• Grok 4.1 Fast (19 November 2025; 2M context; launched with an "Agent Tools API")

These release dates are Wikipedia-only (Wikipedia: Grok). The retired model IDs now redirect to grok-4.3 at $1.25 / $2.50: reasoning IDs map to low effort and non-reasoning IDs to none (xAI retirement notice). An automation still calling grok-4-1-fast-reasoning therefore runs on a different model at a different price, without raising an error. That is the strongest argument for pinning versions. xAI's aliases work like this: <model> points to the latest stable version, <model>-latest to the newest, and <model>-<date> to one fixed release (xAI models and pricing).

Sources for the table: xAI models and pricing, OpenRouter models API, xAI release notes, xAI Grok 4.3 docs page (Hacker News listing) and geotoolbox.ai. The "0309" suffix suggests a March snapshot, but that is inferred from the ID. The grok-4.20 context figures conflict, so test the limit before relying on it.

Since March 2026, xAI has shipped a model roughly every month. Wikipedia reports a deliberate monthly cadence, citing a Musk statement that the docs do not confirm (Wikipedia: Grok). The recent releases:

• Grok 4.7: xAI calls it a new, larger base model, "twice as fast, at half the price" of comparable models (vendor-reported) (xAI: Grok 4.7). Its knowledge cutoff is May 2026 (xAI models and pricing). xAI's release notes confirm text and image input with text-only output, no text output limit, and four reasoning-effort settings: low, medium, high (the default) and xhigh (xAI release notes; also reported by Let's Data Science and VPS Ranking).
• Grok 4.6: its knowledge cutoff is 1 February 2026, and a "faster edition" costs twice the standard price (secondary) (DataNorth AI). For Grok 4.7, xAI's pricing page documents the equivalent "Grok 4.7 Fast" at 2x standard rates (1.5x for long-context requests), sold only inside Cursor and Grok Build (xAI pricing).
• Grok 4.5: it targeted coding and agent work, and Cursor offered it at launch (secondary) (Codersera). It reached EU API users only later in July (xAI release notes). Musk called it "Opus-class" but faster and cheaper, which Wikipedia records as the company's own assessment.

No Grok 4.x weights are public. xAI's Hugging Face organisation lists only grok-1 and grok-2, although only the top five repositories were checked (Hugging Face xai-org listing). For self-hosting, see DeepSeek and Other models.

Pricing rewards short prompts: $2 / $6 under 200K tokens, double above

The table comes from xAI's models page (xAI models and pricing). Every row was re-read on 11 October 2026 from the markdown version of the vendor page (xAI models page), and OpenRouter lists the same rates (OpenRouter models API).

The 200K rule covers the whole request. In xAI's words, "requests whose prompt reaches the listed token threshold are billed at the higher rate for all tokens in the request", output included (xAI models page; corroborated by Let's Data Science). On grok-4.7, a 199,000-token prompt costs about $0.40 in input, while a 201,000-token prompt costs about $0.80. The 1M-context models carry the same step, so switching to them lowers the base price but keeps the surcharge.

Rivals charge less for long prompts:

• OpenAI: 2x input and 1.5x output above 272K tokens, with the whole request repriced (GPT-6.1 Sol $4 / $15) (OpenAI pricing).
• Google: 2x input and 1.5x output above 200K on Gemini 3.1 Pro and 2.5 Pro; input above 200K bills all input and output at the long rate (Gemini API pricing; Google Cloud pricing).
• Anthropic: the full 1M context at standard rates on Claude 4.6 and later, except Haiku 5.5 (Claude Platform pricing).

Among these flagship tiers, only xAI doubles the output price. The exception elsewhere is Anthropic's cheapest model: Haiku 5.5 costs five times as much for input and output once a prompt passes 100K tokens. See AI automation costs.

Caching saves less than at rivals. Grok 4.7's cached input costs 25% of its normal input price. Claude Sonnet 5.5 cache reads cost $0.10 against $2 input since Anthropic's 7 October cut (Claude Platform pricing), as does GPT-6.1 Sol's cached input (OpenAI pricing). That is 5% of the input price. A cached 50K-token system prompt costs $0.025 per call on Grok 4.7 against $0.005 on Sonnet 5.5, so automations that reuse long prefixes lose much of Grok's headline advantage. Grok's real saving is on output: $6 against $10 for Sonnet 5.5 and GPT-6.1 Sol.

Batch discounts are smaller, and absent on the flagships. The Batch API typically returns results within 24 hours, does not count toward per-minute limits, and discounts text models only (xAI Batch API). xAI's pricing page, re-read on 11 October 2026, lists 20% off for grok-4.3 and the three grok-4.20 models and states that "models not listed above have no batch discount", which leaves grok-4.5, 4.6, 4.7 and grok-build-0.1 at full price in batch (xAI pricing). OpenRouter's grok-4.3 batch route at $1 / $2 matches (OpenRouter models API). OpenAI's batch tier, by contrast, is half the standard price (OpenAI pricing). A news snippet cited an OpenRouter route for Grok 4.7 at $1.60 / $4.80 (mer.vin). OpenRouter's own listing shows $2 / $6, and this document uses the listing.

Two other multipliers apply. Priority processing costs 2x on all token types, and the US regional endpoint (us.api.x.ai/v1) costs 1.1x, which is $2.20 / $0.55 / $6.60 for grok-4.7 under 200K tokens (xAI pricing). See the data-handling section for what the US endpoint covers.

Sources: xAI models and pricing and UsagePricing tracker, which calls 2.0 a 60% premium. grok-voice-latest has routed to 2.0 since 5 August 2026 (xAI release notes), so check whether 1.0 is still sold.

Sources: xAI pricing, xAI X search, OpenRouter models API, OpenAI pricing, Claude pricing and Gemini API pricing. xAI's pricing page, re-read on 11 October 2026, gives the unit as cost per 1,000 calls; X search is billed per item returned. Only successful tool calls are billed.

Worked example (arithmetic, no caching). 10,000 extraction jobs a month, each with a 20K-token prompt and a 2K-token answer, cost $520 on grok-4.7 (200M × $2 plus 20M × $6) and $300 on grok-4.3. Long reasoning output widens the gap (see Benchmarks).

The API now covers the whole automation toolkit

The Responses API is xAI's recommended interface. The examples use the OpenAI Python and JavaScript SDKs with base_url="https://api.x.ai/v1", and a migration guide is provided (xAI Responses comparison). The reference still lists POST /v1/messages as "compatible with the Anthropic API". It is marked: "Deprecated: The Anthropic SDK compatibility is fully deprecated. Please migrate to the Responses API or the gRPC API." No end date is given (xAI llms-full.txt). Building on the OpenAI-style interface keeps the cost of switching vendors low. See Agent frameworks.

Function calling follows the usual pattern. Tools are declared as JSON schemas, the model returns a function_call item, and the caller sends back a function_call_output. When streaming, each call arrives whole in a single chunk (xAI function calling).

Structured outputs are documented as guaranteed for supported schema features when response_format.type = "json_schema" is set. Tool arguments always follow the tool's schema, because strict mode is always on. Key details (xAI structured outputs):

• Enforced string formats: date, time, date-time, email, uuid, ipv4, ipv6 and uri.
• Guarantee limits: strings up to 2,048 characters, arrays up to 256 items, objects up to 64 properties.
• Best effort only: if/then/else, not and multi-subschema allOf.
• Defaults and errors: additionalProperties defaults to false, and enums with no variants return HTTP 400.

This is an unusually explicit guarantee for invoice, form and ticket extraction, but test it on your own schemas.

Built-in tools run on xAI's servers: web search, X search, a code interpreter, image generation and collections search (xAI tools overview). Without them, Grok has no access to realtime events (xAI models and pricing).

• Web search takes domain allow and deny lists (at most five allowed domains) and has offered image search since May 2026 (xAI web search; xAI release notes).
• X search has no equivalent among rivals. It filters by up to 20 handles and by date range, and it understands images and video (xAI X search).
• Code execution is a Python sandbox (xAI code execution).
• Collections pair a Files API with a collections_search tool, called file_search in the OpenAI format (xAI collections search).
• Remote MCP tools are supported and billed as tokens (xAI pricing).

Multi-agent research and voice are the specialised options. The beta grok-4.20-multi-agent model coordinates several agents that search and combine findings, and its interface may change (xAI multi-agent). Realtime voice runs over a WebSocket at wss://api.x.ai/v1/realtime, with session updates, server-side voice-activity detection and in-session tools. Browser clients use short-lived "ephemeral tokens" instead of the API key (xAI voice; xAI speech-to-speech). grok-voice-transcribe-1.0 reached end of life on 2 October 2026, and its requests now route to grok-voice-transcribe-2.0 at the same price (xAI release notes).

The docs index also names these features, which were not read in full (xAI llms.txt):

• prompt caching
• context compaction (May 2026)
• priority processing (June 2026)
• a WebSocket mode for the Responses API (May 2026)
• deferred and async requests
• safety_identifier (September 2026), for attributing abuse to individual end users

Three constraints to plan around (xAI models and pricing):

• logprobs is silently ignored on grok-4.20 and newer, so confidence scores must come from elsewhere.
• grok-4.7 returns its reasoning only as reasoning.encrypted_content; xAI says reasoning tokens are "billed as part of your total consumption" without naming the rate (xAI reasoning).
• Image input is limited to JPG and PNG up to 20 MiB.

Rate limits rise with cumulative spend since 1 January 2026. Tiers upgrade automatically and never downgrade, and limits above Tier 4 are available on request as an Enterprise tier (xAI rate limits).

Across the same tiers, grok-4.3, grok-4.20 and grok-build-0.1 get 37 to 208 requests per second. The multi-agent model gets only 9 to 56. Going over a limit returns HTTP 429, and the per-second limit is "derived from your per-minute request budget (RPM / 48, with a minimum of 2)". The RPS and TPM column labels were confirmed on the live page on 11 October 2026 (xAI rate limits). Tier 0 is far above typical SME volumes.

Beyond the API: Grok Build, Grok Bot and business seats

Grok Build is a terminal coding agent. It was open-sourced under Apache-2.0 on 16 July 2026 according to Wikipedia; a Hacker News listing gives 15 July (Wikipedia: Grok; GitHub: grok-build). Wikipedia says it comes with SuperGrok at $30 a month (xAI's pricing page also refers to a Grok Build free tier: xAI pricing) and scored 70.8% on SWE-bench Verified as of 15 May 2026. That is a product score with the harness unstated (Wikipedia-only). Wikipedia also describes a report from 14 July 2026 that Grok Build uploaded whole repositories, secrets included, to SpaceXAI cloud storage even with "Improve the model" switched off. SpaceXAI changed a server setting the same day. This rests on a single chain of sources, but it is reason enough to run coding agents only on repositories that have been scanned for secrets.

Grok Bot offers "durable AI teammates": agents on a persistent cloud computer, with messaging, approvals, connectors and routines. Its beta launched on 11 August 2026 (Wikipedia), and it is now listed as available. The docs index covers team bots, private networks, TLS-inspecting proxies and a security FAQ, none of which were read (xAI Grok Bot overview). It competes for the always-on agent work described in Use cases.

Grok Business and Grok Enterprise are seat-based workspaces. Licences are assigned from console.x.ai, sharing is limited to licensed members, and Enterprise allows custom retention (xAI Grok user guide). SpaceX's prospectus names both as subscription products (SpaceX S-1/A No. 2). Connectors are indexed for Google Drive, Gmail and Calendar, Outlook, SharePoint, OneDrive, Teams, Salesforce, and custom MCP tunnelling (xAI llms.txt). No seat price was found.

Data handling: 30-day retention by default, and zero retention switches features off

"By default, all API requests and responses are stored on our servers (encrypted at rest) for 30 days for auditing purposes... SpaceXAI does not train on this data" (xAI security FAQ). Zero Data Retention (ZDR) applies team-wide, and xAI does not recommend it for most customers. Under ZDR the following are unavailable:

• per-key request logging
• the stateful Responses API
• new Files API uploads
• the Collections API
• the Batch API
• deferred completions
• server-stored image and video outputs
• voice-agent conversation history

A business that needs zero retention therefore loses both hosted RAG and the Batch API, and must run its own vector store. ZDR is self-serve for a team admin "where available" (xAI security FAQ).

The docs say "We are SOC 2 Type 2 compliant", point NDA customers to a Trust Center, and offer a HIPAA BAA on request (vendor-reported; not checked against the Trust Center) (xAI llms-full.txt). Enterprise terms and the DPA are at x.ai/legal (xAI general FAQ).

Residency is US-only and partial. The global endpoint "may route requests between regions for capacity or reliability, so the processing location is not guaranteed". The US regional endpoint (us.api.x.ai/v1) guarantees US request handling, inference, moderation and storage of retained request data, at 1.1x token rates. It serves only grok-4.7 and grok-4.6, and it excludes Files, Collections, server-side tools such as web and X search, and the image, video and voice APIs. xAI warns that it "is not, by itself, a comprehensive data-residency guarantee" (xAI regional endpoints; xAI pricing). No EU or India endpoint is documented. By comparison, OpenAI charges 10% extra on its data-residency endpoints, processes in-region only in the US, EU and UAE, and lists India for storage at rest only (the 10% applies to models released on or after 5 March 2026, and the India option requires Modified Abuse Monitoring or ZDR), while Anthropic charges 1.1x for US-only inference on Claude 4.6 and later (OpenAI data controls; OpenAI pricing; Claude Platform pricing). None of the three direct APIs documents in-country inference in India. See Governance and security.

Where Grok runs: direct API, three hyperscaler catalogues and the router layer

US businesses can pay through their existing Azure, Google Cloud or AWS contracts. For Indian SMEs, direct USD billing is the simplest route, since no INR price list exists. Grok has little share of routed developer traffic. In OpenRouter's request-share chart for the week of 28 September 2026, xAI is not among the nine named authors; the smallest of them, Mistral, had 2.0% (OpenRouter rankings). That chart measures requests, not enterprise spend. See Companies.

Benchmarks put Grok 4.7 in the upper tier, about 11 points off the lead

Artificial Analysis (AA) scores Grok 4.7 (xhigh) at 46 on its Intelligence Index, "well above average" against a median of 26. It runs at 81 tokens per second and costs $3.74 per index task (Artificial Analysis: Grok 4.7). Grok 4.6 (high) scores 44 at 61 tokens per second and $1.48 per task (Artificial Analysis: Grok 4.6). The gap in cost is verbosity: Grok 4.7 used 240M output tokens on the index, against 94M for Grok 4.6 and a median of 82M. Wikipedia reports the same pattern (Wikipedia: Grok). For automation, budget per task rather than per token.

One widely repeated figure is wrong. DataNorth AI and a Hacker News listing put Grok 4.6 at 61 on the AA index (DataNorth AI; xAI Grok 4.6 post (Hacker News listing)). AA's own page says 44, and 61 is that model's output speed, so this document uses 44. Claims about speed also conflict. A snippet said Grok 4.7 matches Grok 4.6's speed (Let's Data Science), but AA measures 81 against 61 tokens per second. AA's 81 is exactly the average, which does not support xAI's "twice as fast" claim.

Sources: Artificial Analysis leaderboard, OpenRouter models API, Moonshot pricing, Z.ai pricing, Alibaba Cloud Model Studio pricing, Anthropic: Claude Opus 5.5, Anthropic: Claude Sonnet 5.5, Anthropic: Claude Haiku 5.5 and OpenAI pricing.

Three comparisons matter:

• Sonnet 5.5 and GPT-6.1 Sol cost the same $2 for input and score about ten and five points higher, though their output price is 67% higher.
• Muse Spark 1.3 is cheaper on both input and output and scores about two points higher, so Grok does not lead on price-performance either.
• Qwen3.8 Max matches Grok's $2 / $6 exactly and scores one point lower, making it the closest substitute.

Other evidence is thin:

• Humanity's Last Exam: Wikipedia lists Grok 4.6 at 42.9, against Kimi K3 at 46.9 and Qwen 3.8 at 42.4. The tool setting was not checked (Wikipedia: Humanity's Last Exam).
• LMArena (arena.ai): on the 8 October 2026 text snapshot (preference Elo), Grok 4.7 ranks #90 at 1443, against a preliminary #1 Gemini 4 Argon (1525) and #2 Claude Opus 5.5 (1507) (LMArena text); no Grok model appears in the "Top 10 Agents" list fetched on 11 October (arena.ai leaderboard).
• Prediction market: a Manifold snapshot gave xAI about 3% odds of holding the top text model on 31 October 2026. That is a forecast, not a measurement (Manifold market).
• xAI's own table: it compares Grok 4.7 at maximum effort with Grok 4.6 at a lower effort on most rows (Let's Data Science).

Two independent boards put Grok 4.7 well below the leaders. On the tbench.ai Terminal-Bench 4.0 board, Grok 4.7 (xhigh, Grok Build harness) scores 37.58%, against 64.85% for Claude Opus 5.5 and 58.18% for GPT-6 Astra (tbench.ai). On Epoch's own FrontierMath runs (Tiers 1-3, indicative) it scores 53.0%, against 91.2% for Opus 5.5 and 93.7% for GPT-6.1 Sol (Epoch benchmark data). ARC-AGI and model-level SWE-bench values for Grok were not found. For the rivals' figures, see Claude, OpenAI and Gemini.

Scale and solvency: a $3.2 billion AI segment losing $6.4 billion

SpaceX's prospectus amendment (S-1/A No. 2, filed 3 June 2026) is the only primary source on Grok's business (SpaceX S-1/A No. 2).

The excerpts read do not say whether X's advertising sits inside this segment. The filing counts about 550 million combined Grok-or-X monthly users at 31 March 2026, of whom about 117 million used Grok's AI features, up from 89 million at 31 December 2025. These are company-defined metrics. They replace a secondary estimate of 35 million users in December 2025. Secondary coverage also reports about 1.9 million SuperGrok subscriptions, a figure not confirmed in the filing. The filing lists the Irish Data Protection Commission's "large-scale inquiry", opened in February 2026, as a risk factor.

The ownership structure changed four times in fourteen months (Wikipedia-only) (Wikipedia: SpaceXAI):

• December 2024: a $6B private round took total funding above $12B.
• March 2025: xAI bought X, valuing X at $33B and xAI at $80B.
• July 2025: Morgan Stanley arranged $5B of debt and $5B of equity, with SpaceX committing $2B.
• February 2026: SpaceX acquired xAI.

SpaceX's filings state no dollar value for the deal. The prospectus (424B4, 12 June 2026) and the Q2 2026 10-Q give only the fixed ratio of 0.1433 SpaceX shares per xAI share (pre-split) and a $75.46-per-share cash election for eligible service providers (SpaceX 424B4). The values in circulation come from the press and Wikipedia, not from a filing: a combined value of about $1.25 trillion (CNBC, snippet), $250B for xAI (Wikipedia's SpaceXAI article), $125B (its IPO article, Wikipedia: SpaceX IPO) and $219B to $294B (bank documents seen by CNBC). Musk cited space-based data centres as the reason for the merger, while analysts pointed to xAI's cash burn (secondary) (D&O Diary; Medium/AIMonks).

SpaceX listed on Nasdaq (SPCX) on 12 June 2026 (Wikipedia-only):

• Pricing: $135 a share, an implied $1.77T valuation. A March preview had expected up to $50B at about $1.75T (Motley Fool).
• First day: closed at $161, up 19%, a market value of about $2.1T.
• Proceeds: $86B against a $75B target.
• Afterwards: shares were down about 50% from their peak by late July 2026.
• Control: Musk holds about 42% of the shares and 82% of the votes.

Cursor now belongs to the group. The S-1/A records an April 2026 option to buy Cursor at an implied $60.0B, with a $1.5B termination fee and an $8.5B deferred services fee under specified conditions. SpaceX signed the merger agreement on 16 June 2026 (8-K), and the merger became effective on 14 August 2026 (8-K, Item 2.01), with 389,289,254 Class A shares issued to Cursor holders (SpaceX 8-K).

Compute rentals to rivals are large (Wikipedia-only):

• Anthropic: contracted on 6 May 2026 for $1.25B a month for all of Colossus 1 (about 220,000 Nvidia GPUs).
• Google: agreed on 5 June 2026 to pay $920M a month for about 110,000 GPUs from October 2026 to June 2029 (CNBC headline).

Together that is about $26B a year, roughly eight times the AI segment's 2025 revenue. Which segment books it is unknown. Renting GPUs to Grok's competitors may soon matter more to the group than Grok itself.

The data centres are in Memphis and Southaven (Wikipedia-only) (Wikipedia: Colossus):

• Colossus 1 started with about 100,000 H100 GPUs, was built in 122 days and went live in 2024.
• Colossus 2's first cluster of about 110,000 GB200-class GPUs and about 210 MW came online around January 2026.
• Power comes from a gas-turbine plant (about 1.2 GW permitted) and Tesla Megapacks.
• Expansion: a third building is meant to take capacity to "nearly 2 gigawatts". Musk's target is at least 1 million GPUs.

A blog's figure of 2 GW and 555,000 GPUs contradicts its own table and is not used (Introl). A $659M expansion permit is reported only by a secondary blog (Basenor). The NAACP sued in April 2026 over the gas turbines, and Reuters reported 59 turbines at Colossus 2 running without federal air permits (Wikipedia-only). See Hardware.

Leadership churn is heavy. Wikipedia records a February 2026 restructuring into four development teams. xAI was founded by Musk and 11 others; about half the co-founders had left by February 2026, and by late March 2026 Musk was the only one remaining (Wikipedia citing Reuters; the late-March point is Wikipedia-only) (Wikipedia: SpaceXAI). Other reports count six exits by February (eWeek; Gigazine). CFO Anthony Armstrong left on 10 April 2026. See People.

From DOGE to GenAI.mil: Grok's US government footprint

US government use has grown in steps:

• 2025, DOGE: DOGE used Grok "heavily" (Reuters, 8 April 2025, via Wikipedia) (Wikipedia: Grok).
• 14 July 2025, Defense Department: a contract with a $200M ceiling through the Chief Digital and AI Office (CDAO). OpenAI, Anthropic and Google got equivalent deals (CBS News; Interesting Engineering). xAI launched "Grok for Government" on Grok 4 at the same time (Maginative). Coverage cited Grok's antisemitic outputs of the week before (ABC News 4).

The 2026 steps are Wikipedia-only (Wikipedia: Grok deepfake scandal):

• 12 January 2026: Defense Secretary Pete Hegseth announced Grok's integration into classified and unclassified DoD networks.
• June 2026: court filings reported that a government Grok was used in the Maven Smart System against Iran. The CDAO said the system supported more than 2,000 munitions against 2,000 targets in 96 hours.
• August 2026: "Grok for Government" (Starshield AI) joined GenAI.mil, accredited for controlled unclassified information at IL5.
• September 2026: customised ChatGPT and Grok were made available on GenAI.mil.

GSA's own release sets the terms. Under the OneGov agreement announced on 25 September 2025, agencies get Grok 4 and Grok 4 Fast for $0.42 per organisation (agency) for 18 months, valid until March 2027, with an upgrade path to FedRAMP and DoD IL-aligned enterprise subscriptions (GSA). The "per user" and "per instance" versions should not be used. Adoption by non-US governments was not researched. See Americas and Europe.

The deepfake scandal made Grok the most-restricted frontier chatbot

Grok had a record of controversy before 2026 (Wikipedia-only) (Wikipedia: Grok; Wikipedia: 2025 in AI):

• May 2025: a "white genocide" output, blamed on an "unauthorized modification".
• 8 July 2025: antisemitic "MechaHitler" replies, which xAI attributed to an upstream code change.
• July 2025: a court-ordered block in Turkey over insults to Erdoğan and Atatürk.
• July to August 2025: "Spicy" modes in Grok Imagine.

Then, from late December 2025, Grok was used on X to make sexualised edits of real people's photos, including minors. On 9 January 2026, xAI limited image generation to paid subscribers (News Arena India), and UK officials called the move "insulting". The table below draws on the Wikipedia scandal article, which cites Reuters, the Guardian, RTÉ/PA and BBC (not opened), unless another source is noted (Wikipedia: Grok deepfake scandal).

The other sources for the table are:

• Wikipedia: 2026 in AI for the block dates and the Baltimore suit.
• Social Samosa and Yahoo News for the French and Malaysian inquiries.
• Anadolu Agency (headline only) and EDRi (search summary) for the EU proceeding under DSA Arts 34, 35 and 42.

Separately, xAI sued Colorado's Attorney General on 9 April 2026 over SB 24-205 (X.AI LLC v. Weiser, D. Colo., No. 1:26-cv-01515), arguing a First Amendment violation (Wikipedia: Colorado AI Act). The United States intervened as a plaintiff on 24 April, and a minute order of 27 April (ECF 24) bars the AG from beginning enforcement under SB 24-205, or legislation replacing or amending it enacted that session, until 14 days after a ruling on xAI's preliminary-injunction motion, which is due within 28 days after final implementing rules; deadlines are stayed. Whether the stay reaches the re-enacted SB 26-189 is arguable from the order's wording, not confirmed (CourtListener docket). xAI is also in the 8th Circuit against Minnesota (SpaceXAI LLC / X.AI LLC v. Ellison, No. 26-2806): an emergency motion for an injunction pending appeal, an opposition filed 21 September 2026 and amicus briefs of 17 and 29 September (filings only; the district-court ruling was not read) (CourtListener). The national blocks lasted only 5 to 22 days and hit the consumer chatbot, not the API. They still showed that South-East Asian regulators can switch Grok off nationally at a few days' notice. See Asia-Pacific and Countries.

India: a notice, a reply and no recorded outcome

On 2 January 2026, MeitY served X Corp with a notice over Grok being used to make derogatory synthetic content about women and to strip clothing from images. The notice demanded an action-taken report within 72 hours and warned that X could lose its intermediary safe-harbour protection. MeitY also reportedly ordered a "comprehensive review" of the Grok ecosystem (The Logical Indian; exchange4media; Digital Policy Alert). Government sources said X's reply set out its policies but did not list specific takedowns or preventive measures (secondary) (CXO Today). A second notice followed on 8 January (reported), and MP Priyanka Chaturvedi filed a complaint. No published action-taken report or enforcement decision was found, although around 11 January "government sources" said about 3,500 items had been blocked and about 600 accounts deleted (secondary, several outlets) (Storyboard18). The safe-harbour lever applies to X as an intermediary. The IT Rules amendment on synthetically generated information (G.S.R. 120(E), notified 10 Feb, in force 20 Feb 2026; secondary) requires a prominent label and cuts takedown to 3 hours for court or government orders and 2 hours for complaints about non-consensual intimate imagery, including deepfake nudity (iPleaders; Gala Law). Whether businesses that build on the Grok API carry obligations under the IT Rules 2021 was not researched, so confirm it with counsel. See India and Governance and security.

What this means for your business

Grok is a competent, reasonably priced second-tier frontier model with one unique asset (live X data), cheap built-in tools and a heavier risk profile than its peers. It fits workloads that use X data or web search heavily, keep prompts under 200K tokens and never produce images of real people, rather than serving as the default for everything.

What it costs per task

Tier 0 gives 150 requests per second with no spend. A prompt of 200K tokens or more doubles the price of the whole request, so chunk below it; cached input costs 25% of input (5% at some rivals); and the 20% batch discount covers only grok-4.3 and grok-4.20. At list price, 10,000 extraction jobs a month (20K-token prompt, 2K-token answer) cost $520 on grok-4.7 and $300 on grok-4.3. Grok 4.7's verbosity puts it at $3.74 per AA index task against $1.48 for Grok 4.6, so budget per task and use the lowest reasoning effort that passes your evaluations. With no INR price list, budget in USD with an exchange-rate buffer.

Data handling and India processing

• Retention: 30 days by default, with no training on API data (xAI). Zero retention is team-wide and switches off new Files uploads, Collections, Batch and the stateful Responses API, so zero-retention RAG needs your own vector store. Record retention in your data-processing terms.
• India and the EU: no endpoint is documented; Bedrock serves Grok from India only through its global profile, and OCI Hyderabad has none. Ask counsel whether building on the Grok API brings IT Rules obligations (not researched). With European users, the EU and UK cases matter; xAI signed only the Safety and Security Chapter of the EU's GPAI Code of Practice (European Commission).
• US: the US endpoint (1.1x) covers only grok-4.7 and grok-4.6 and excludes files, collections and server-side tools, so a US-only design must avoid them or confirm with xAI. Foundry and Vertex billing (vendor-claimed), a HIPAA BAA on request (vendor-reported), the GSA OneGov deal (valid until March 2027) and GenAI.mil help public-sector buyers.

Risks and controls

• Pin dated model IDs. Old IDs were silently redirected in May 2026, and xAI now ships monthly.
• Pass safety_identifier to attribute abuse to end users, and use ephemeral tokens for browser voice.
• Build on the OpenAI-compatible Responses API so changing vendor means little more than changing base_url, with a second model behind it (Tool stack, Playbook).
• Treat vendor continuity as a standing risk. The AI segment lost $6.4B in 2025 and most founders have left; track MeitY, Ofcom and EU releases too.

What to test before relying on it

Test the structured-output guarantee on your own schemas, the grok-4.20 context limit (1M or 2M), latency against xAI's "twice as fast" claim, Indian-language quality for text and voice (not researched), output tokens at each reasoning effort, and the reasoning-token rate on a test bill. Read the DPA and request the SOC 2 report before sending regulated data.

Questions to ask any vendor or implementation partner

• Which dated model IDs and which endpoint does each workflow use, and how is a silent redirect caught?
• Which tools fall outside the US guarantee or zero retention, and what replaces them?
• How are prompts kept under 200K tokens, and which tested fallback sits behind the same interface?

See also Future outlook, the Timeline, the Glossary and the index.

Open questions and unverified items

Sources

1. Wikipedia: SpaceXAI
2. xAI docs: models page (developers)
3. Wikipedia: Grok (chatbot)
4. xAI docs: 15 May 2026 retirement notice
5. xAI docs: models and pricing
6. OpenRouter models API
7. xAI docs: release notes
8. xAI docs: Grok 4.3 model page (Hacker News listing)
9. geotoolbox.ai: Grok 5
10. xAI: Grok 4.7 announcement
11. Let's Data Science: xAI releases Grok 4.7
12. VPS Ranking: xAI Grok 4.7
13. DataNorth AI: xAI releases Grok 4.6
14. Codersera: Grok 4.5 launch guide
15. Hugging Face: xai-org model listing
16. OpenAI API pricing
17. Google Gemini API pricing
18. Claude Platform pricing docs
19. Anthropic: Claude Haiku 5.5
20. xAI docs: Batch API
21. xAI docs: pricing
22. mer.vin: Grok 4.7 arrives with the same price and a hidden pricing tier
23. UsagePricing: xAI price change tracker, 4 Aug 2026
24. xAI docs: X search
25. Claude pricing
26. xAI docs: Responses API comparison and migration
27. xAI docs: llms-full.txt
28. xAI docs: function calling
29. xAI docs: structured outputs
30. xAI docs: tools overview
31. xAI docs: web search
32. xAI docs: code execution
33. xAI docs: collections search
34. xAI docs: multi-agent
35. xAI docs: voice
36. xAI docs: speech-to-speech
37. xAI docs: llms.txt index
38. xAI docs: rate limits
39. GitHub: xai-org/grok-build
40. xAI docs: Grok Bot overview
41. xAI docs: Grok user guide
42. SpaceX S-1/A No. 2 (SEC EDGAR)
43. xAI docs: security FAQ
44. xAI docs: general FAQ
45. xAI docs: Microsoft Foundry guide
46. xAI docs: Google Cloud Vertex AI guide
47. GitHub Docs: supported AI models in GitHub Copilot
48. Wikipedia: Telegram (software)
49. Wikipedia: Optimus (robot)
50. OpenRouter rankings
51. Artificial Analysis: Grok 4.7
52. Artificial Analysis: Grok 4.6
53. xAI: Grok 4.6 announcement (Hacker News listing)
54. Anthropic: Claude Opus 5.5
55. Anthropic: Claude Sonnet 5.5
56. Wikipedia: Humanity's Last Exam
57. arena.ai leaderboard
58. Manifold: which company will have the #1 model
59. CNBC: Musk's xAI and SpaceX, biggest merger ever (3 Feb 2026)
60. Wikipedia: Initial public offering of SpaceX
61. Motley Fool: SpaceX absorbed xAI at a combined $1.25 trillion valuation (31 Mar 2026)
62. D&O Diary: the SpaceX-xAI merger
63. Medium/AIMonks: why SpaceX and xAI merged in 2026
64. CNBC: Google to pay SpaceX $920M a month for xAI compute (5 Jun 2026)
65. Wikipedia: Colossus (data center)
66. Introl: xAI Colossus 2 gigawatt expansion (Jan 2026)
67. Basenor: xAI files $659M permit to expand Colossus 2
68. eWeek: xAI founder departures and Musk reorganization
69. Gigazine: xAI co-founders leaving (12 Feb 2026)
70. CBS News: xAI Pentagon contract
71. Interesting Engineering: xAI Grok Pentagon contract
72. Maginative: xAI launches Grok for Government
73. ABC News 4: Pentagon to start using Grok amid controversy
74. Wikipedia: Grok sexual deepfake scandal
75. Wikipedia: 2025 in artificial intelligence
76. News Arena India: Grok restricts image generation after deepfake backlash
77. Wikipedia: 2026 in artificial intelligence
78. Social Samosa: France and Malaysia join India in condemning Grok deepfakes
79. Yahoo News: French and Malaysian authorities investigating Grok
80. Anadolu Agency: EU opens formal probe into Grok
81. EDRi: EU probes X's Grok over AI-generated harm
82. Wikipedia: Colorado AI Act
83. The Logical Indian: MeitY issues notice to X over Grok deepfakes
84. exchange4media: X gets MeitY notice over Grok misuse
85. Digital Policy Alert: MeitY investigation into X over Grok
86. CXO Today: the Grok AI controversy
87. Wikipedia: Elon Musk
88. Vayam: the global crisis of AI harassment in 2026
89. xAI docs: regional endpoints
90. OpenAI docs: Your data (data retention and residency controls)
91. SpaceX prospectus, 424B4 (12 Jun 2026)
92. SpaceX Form 10-Q, Q2 2026
93. SpaceX Form 8-K (14 Aug 2026), Cursor merger effective
94. GSA: GSA and xAI partner to accelerate federal AI adoption (25 Sep 2025)
95. CourtListener: X.AI LLC v. Weiser docket (D. Colo.)
96. CourtListener: SpaceXAI LLC v. Ellison docket (8th Cir.)
97. Artificial Analysis: Models leaderboard
98. Moonshot AI: Kimi API pricing
99. Z.ai: API pricing
100. Alibaba Cloud Model Studio: Model pricing
101. LMArena: Text leaderboard
102. tbench.ai: Terminal-Bench 4.0 board
103. Epoch AI: Benchmark data (bulk download)
104. AWS: Amazon Bedrock model support by Region
105. Oracle: OCI Generative AI model endpoint regions
106. xAI docs: reasoning
107. Google Cloud: Gemini Enterprise Agent Platform pricing
108. Storyboard18: after MeitY notice, X takes down Grok content and deletes 600 accounts (secondary)
109. iPleaders: IT Rules 2026 amendment (secondary)
110. Gala Law: IT Rules amendment on synthetic content (secondary)
111. European Commission: GPAI Code of Practice