Grok is the model family from xAI, which now appears as SpaceXSI in its API documentation (previously SpaceXAI); the API and console still run under x.ai. The current reference models are Grok 4.7 (since 21 September 2026) as the flagship for coding, agentic tasks and knowledge work, the lower-priced all-rounder Grok 4.3 with 1 million tokens of context and the coding model Grok Build 0.1. In addition there are the image and video models of Grok Imagine and the Grok Voice API.
Enterprises in the EU can reach Grok via the xAI API, Microsoft Foundry, Amazon Bedrock and the Gemini Enterprise Agent Platform on Google Cloud. None of these routes offers documented EU data residency; the details are in the section on EU availability.
innFactory AI Consulting from Rosenheim, Germany advises enterprises in the DACH region on the selection and GDPR-compliant integration of AI models, including via CompanyGPT. This overview reflects the state as of October 2026.
Current models at a glance
| Model | API ID | Release (API) | Context | Price input / output per 1M tokens | Focus |
|---|---|---|---|---|---|
| Grok 4.7 | grok-4.7 | 21 September 2026 | 500k | $2 / $6 | Flagship: coding, agents, knowledge work |
| Grok 4.6 | grok-4.6 | 12 August 2026 | 500k | $2 / $6 | Predecessor; also In-Region (US) on Bedrock |
| Grok 4.5 | grok-4.5 | 8 July 2026 | 500k | $2 / $6 | superseded by Grok 4.6 and 4.7 |
| Grok 4.3 | grok-4.3 | May 2026 | 1M | $1.25 / $2.50 | low-cost all-rounder, 1M context |
| Grok 4.20 | grok-4.20-0309-reasoning, grok-4.20-0309-non-reasoning, grok-4.20-multi-agent-0309 | 10 March 2026 | 1M | $1.25 / $2.50 | reasoning, non-reasoning, multi-agent research (beta) |
| Grok Build 0.1 | grok-build-0.1 | 29 May 2026 | 256k | $1 / $2 | agentic coding |
Prices apply to prompts under 200,000 tokens; above that the rate doubles (see Pricing).
Grok 4.7
Per xAI, Grok 4.7 is the model “for code and everything else” and is based on a new, larger base model than Grok 4.6. It was trained with a longer reinforcement-learning run on tasks that take many hours. Per the provider, Grok 4.7 checks its own work more thoroughly, copes better with long contexts and was trained on the agent harness of Grok Bot. Price and speed match Grok 4.6.
Key data per the xAI documentation: 500,000 tokens of context, text and image input, text output without an output limit, reasoning effort low, medium, high (default) and xhigh, knowledge cutoff May 2026. Function calling, structured outputs, the server-side tools web search, X search and code execution, and the Batch API with a 20% discount are supported. The variant Grok 4.7 Fast is the same model on faster infrastructure, at double the price and available only in Cursor and Grok Build, not via the public API.
For Grok 4.7 (xhigh) versus Grok 4.6 (high), xAI cites, among others, 46.3% instead of 40.4% on CursorBench 4.0, 37.6% instead of 20.3% on Terminal-Bench 4.0, 64.0% instead of 53.0% on EEBench and 56.7% instead of 48.5% on HealthBench Professional. On DeepSWE v1.1, xAI gives 71.0% at high effort, and 1695 Elo on GDPval (Grok 4.6: 1605).
Grok 4.6
Grok 4.6 (12 August 2026) has the same price, the same context window and the same reasoning levels as Grok 4.7, but an older knowledge cutoff (1 February 2026). Per xAI, the focus is on long-running agents and interactive and visual work. xAI cites 61 points on the Artificial Analysis Intelligence Index (Grok 4.5: 56), 69.9% on CursorBench 3.2, 65.9% on DeepSWE 1.1 and 61.3% on FrontierCode 1.1 (Extended).
Grok 4.6 remains relevant mainly on Amazon Bedrock: there, Grok 4.6 additionally has In-Region inference via bedrock-mantle in us-west-2, which Grok 4.7 does not. Grok 4.6 is also available in GitHub Copilot.
Grok 4.5
Grok 4.5 has been available on the API since 8 July 2026 and, per the release notes, also for EU users in the API console since 17 July 2026. It was the first model with 500,000 tokens of context, offers the reasoning levels low, medium and high and, at $0.30 per 1 million tokens, has the lowest cache-read price of the 500k models. xAI cites 83.3% on Terminal-Bench 2.1 and 62.0% on DeepSWE 1.0, with an average of 15,954 output tokens per SWE-Bench Pro task. Grok 4.5 is not listed on Microsoft Foundry, Amazon Bedrock or Google Cloud; the successor at the same list price is Grok 4.7.
Grok 4.3 and Grok 4.20
Grok 4.3 (May 2026) is the family’s low-cost all-rounder: 1 million tokens of context, text and image input and reasoning effort from none to xhigh (default low). Via the Batch API, xAI grants a 20% discount for Grok 4.3 and Grok 4.20, as for Grok 4.7, 4.6 and 4.5. Since 15 May 2026, Grok 4.3 is also the redirect target for the shut-down older models. Per the model page, xAI runs Grok 4.3 in the regions us-east-1, eu-west-1 and us-west-2.
Grok 4.20 (since 10 March 2026) is available as a reasoning and a non-reasoning model and as the multi-agent model grok-4.20-multi-agent, in which 4 or 16 agents research in parallel and a leader agent merges the answer. Per xAI, the multi-agent API is in beta. Price and context match Grok 4.3.
Grok Build 0.1
Grok Build 0.1 has been available as a public beta on the API since 29 May 2026 and is trained specifically for agentic coding, such as web development, debugging and MCP use. The model has 256,000 tokens of context, processes text and images and delivers over 100 tokens per second per xAI, at $1 input and $2 output per 1 million tokens. The earlier ID grok-code-fast-1 is now an alias of grok-build-0.1. The default model in the Grok Build coding agent is now Grok 4.7.
Image, video and voice
Besides the language models, xAI offers Grok Imagine Image 2.0 for image generation and editing, Grok Imagine Video 1.5 for text, image and reference-to-video up to 1080p, and the Grok Voice API with speech-to-speech (grok-voice-3.0, since 14 October 2026), text-to-speech and speech-to-text (grok-voice-transcribe-2.0, since 17 September 2026).
Pricing
xAI API prices, each input / cache read / output per 1 million tokens. If the prompt reaches 200,000 tokens, the higher rate applies to all tokens of the request.
| Model | < 200k tokens | from 200k tokens | Context | Batch discount |
|---|---|---|---|---|
| Grok 4.7 | $2.00 / $0.50 / $6.00 | $4.00 / $1.00 / $12.00 | 500k | 20% |
| Grok 4.6 | $2.00 / $0.50 / $6.00 | $4.00 / $1.00 / $12.00 | 500k | 20% |
| Grok 4.5 | $2.00 / $0.30 / $6.00 | $4.00 / $0.60 / $12.00 | 500k | 20% |
| Grok 4.3 | $1.25 / $0.20 / $2.50 | $2.50 / $0.40 / $5.00 | 1M | 20% |
| Grok 4.20 (reasoning, non-reasoning, multi-agent) | $1.25 / $0.20 / $2.50 | $2.50 / $0.40 / $5.00 | 1M | 20% |
| Grok Build 0.1 | $1.00 / $0.20 / $2.00 | $2.00 / $0.40 / $4.00 | 256k | – |
Prices per provider.
Surcharges and variants:
- US endpoint (
us.api.x.ai/v1, Grok 4.7 and 4.6 only): 1.1 times the price, so $2.20 / $0.55 / $6.60 for Grok 4.7 - Priority processing: double the price for all token types
- Grok 4.7 Fast (Cursor and Grok Build only): $4 / $1 / $12 under 200k tokens, $6 / $1.50 / $18 above
Server-side tools (per 1,000 calls): web search and code execution $5 each, collections search $2.50; X search is billed per hit ($5 per 1,000 posts, $10 per 1,000 profiles).
Image, video, voice: Grok Imagine Image 2.0 costs from $0.04 per image (1K, quality low), Imagine Video 1.5 from $0.08 per second (480p; Lite variant from $0.02). Speech-to-speech costs $0.08 per minute, speech-to-text $0.10 per hour (REST) or $0.20 (streaming), text-to-speech $15 per 1 million characters.
Amazon Bedrock (standard tier, input / output / cache read per 1 million tokens): Grok 4.7 and Grok 4.6 cost $2.00 / $6.00 / $0.50 via the Global profile and $2.20 / $6.60 / $0.55 via the US Geo profile; Grok 4.6 In-Region likewise $2.20 / $6.60 / $0.55. Grok 4.3 costs $1.25 / $2.50 / $0.20 In-Region. Per AWS, priority is billed at 1.75 times and flex at 0.5 times; there is no reserved tier.
EU availability and GDPR
xAI API
The direct API has two access paths. The global endpoint api.x.ai may, per xAI, route requests between regions for capacity or reliability reasons; the processing location is not guaranteed. The US endpoint us.api.x.ai/v1 guarantees that request handling, inference, moderation and stored request data remain in the US; it serves only Grok 4.7 and Grok 4.6. Files, collections and server-side tools such as web search or code execution are excluded from this guarantee. There is no EU endpoint. xAI itself points out that a regional endpoint alone is not a comprehensive data residency commitment.
The model pages list the xAI regions for Grok 4.7 and 4.6 as us-east-1, us-west-2 and us-central-1, and for Grok 4.3 additionally eu-west-1. EU processing cannot be derived from this as long as no EU endpoint exists.
On data processing: xAI stores API requests and responses encrypted for 30 days by default for abuse checks and, by its own statement, does not train on them without explicit permission. Zero Data Retention can be enabled by team admins in the console; however, the stateful Responses API, files, collections, the Batch API and deferred completions, among others, cannot then be used. xAI states that it is SOC 2 Type 2 certified.
Microsoft Foundry
The Foundry catalog includes grok-4.7 (GA), grok-4.6, grok-4.3, grok-4-20-reasoning and grok-4-20-non-reasoning (each preview), grok-4.1-fast-reasoning and -non-reasoning, as well as grok-4 and grok-code-fast-1 (both with registration). Grok 4.5 is not listed. For generally available Grok models from Grok 4.7 onwards, Microsoft commits to a minimum availability of six months. The limits partly differ from the xAI API: grok-4.6 has 200,000 tokens of context on Foundry, grok-4.3 200,000 input tokens and 8,192 output tokens.
Per the region table, the current Grok models can be selected in the deployment type “Global Standard” in twelve European regions: France Central, Germany West Central, Italy North, Norway East, Poland Central, Spain Central, Sweden Central, Switzerland North, Switzerland West, UK South, UK West and West Europe (UK South and UK West are in the United Kingdom, Norway and Switzerland outside the EU). Per Microsoft, “Global Standard” means that prompts and responses may be processed in any Azure region. Microsoft offers “Data Zone Standard” for Grok only in the US data zone, not in the EU; “Standard/Regional” does not exist for Grok.
Amazon Bedrock
| Model | On Bedrock since | Access | EU data residency |
|---|---|---|---|
| Grok 4.7 | 28 September 2026 | us.xai.grok-4.7 (US Geo), global.xai.grok-4.7; from EU regions only Global | No |
| Grok 4.6 | 18 August 2026 (GA 19 August) | us.xai.grok-4.6, global.xai.grok-4.6; In-Region only via bedrock-mantle in us-west-2 and GovCloud | No |
| Grok 4.3 | 15 June 2026 | In-Region via bedrock-mantle in us-west-2, us-east-1, us-east-2, us-gov-west-1 | No |
Per the AWS model cards, no EU geo inference profile exists for any Grok model. From Frankfurt, Ireland, Paris, Milan, Spain, Stockholm or Zurich, Grok is reachable only via the globally routing Global profile. On Bedrock, Grok 4.7 supports guardrails, reasoning, structured outputs and implicit prompt caching, but no server-side tool use. AWS commits to operating Grok 4.7 until at least 28 September 2027.
Google Cloud (Gemini Enterprise Agent Platform)
On the Gemini Enterprise Agent Platform (formerly Vertex AI), Grok models are available as partner models via Model Garden. Grok 4.6 (since 21 August 2026) and Grok 4.7 (release date 30 September 2026) are GA there; both run in the US multi-region and via the global endpoint, and per Google ML processing takes place in the US. Grok 4.3 (preview) and Grok 4.20 (GA) are available only via the global endpoint; Grok 4.1 Fast is no longer listed. Grok 4.5 is not offered. There is no EU region for Grok on Google Cloud.
Self-hosting
Only older models can be self-hosted: Grok-1 (March 2024) under Apache 2.0 and Grok 2, whose weights xAI published on Hugging Face on 22 August 2025. Grok 2 is under the xAI Community License: commercial use is permitted if xAI’s Acceptable Use Policy is observed; the model and its outputs may not be used to train other base models, and products must display “Powered by xAI”. For operation, xAI cites around 500 GB of weights and 8 GPUs with over 40 GB of memory each. The current API models are not available as open weights.
Summary
| Route | Current Grok models | EU data residency |
|---|---|---|
| xAI API | all | No (global endpoint without guarantee, US endpoint only) |
| Microsoft Foundry | Grok 4.7, 4.6, 4.3, 4.20 | No (“Global Standard”, no EU data zone) |
| Amazon Bedrock | Grok 4.7, 4.6, 4.3 | No (no EU geo profile) |
| Google Cloud | Grok 4.7, 4.6, 4.3, 4.20 | No (US and global only) |
| Self-hosting | Grok-1, Grok 2 | Yes, but older model generation |
Migration and breaking changes
- Responses API: xAI recommends the Responses API and treats Chat Completions as a legacy interface. On the Responses API, Grok 4.7 always returns
reasoning.encrypted_content; these reasoning items should be passed back unchanged in follow-up requests. - Reasoning effort:
xhighexists from Grok 4.6; on Grok 4.5 it is treated ashigh. Reasoning cannot be switched off on Grok 4.7 and 4.6; if you need answers without reasoning, use Grok 4.3 withnone. - logprobs: From Grok 4.20,
logprobsandtop_logprobsare silently ignored. - Prompt caching: For Grok 4.7, xAI recommends a
prompt_cache_key(or the headerx-grok-conv-idfor Chat Completions) so that cache hits work reliably. - Model aliases:
<model>points to the latest stable version,<model>-latestto the latest version,<model>-<date>stays fixed. For reproducible results, use the dated ID. - Platform differences: Context and output limits on Foundry are partly smaller than on the xAI API; on Bedrock, Grok 4.7 and 4.6 are addressed via inference profiles (
us.orglobal.) instead of the plain model ID.
Model lifecycle
On 15 May 2026, xAI removed several older models from the API: grok-4-1-fast-reasoning, grok-4-1-fast-non-reasoning, grok-4-fast-reasoning, grok-4-fast-non-reasoning, grok-4-0709, grok-code-fast-1 and grok-3. The IDs keep working, but requests run on grok-4.3 (reasoning models with low, non-reasoning models with none) and are billed at its price; grok-code-fast-1 runs on grok-build-0.1.
The following are therefore superseded on the xAI API:
| Model | Release | Characteristic | Successor |
|---|---|---|---|
| Grok 4.1 Fast | 19 November 2025 | 2 million tokens of context | Grok 4.3 |
Grok 4 (grok-4-0709) | 9 July 2025 | 256,000 tokens of context | Grok 4.3 |
| Grok Code Fast 1 | 2025 | Coding model | Grok Build 0.1 |
| Grok 3 | February 2025 (API from April 2025) | – | Grok 4.3 |
| Grok 4.3 Beta | 17 April 2026 | Preview version | Grok 4.3 (GA) |
On Microsoft Foundry, grok-4.1-fast, grok-4 and grok-code-fast-1 are still listed; on Google Cloud, Grok 4.1 Fast is no longer in the catalog. Grok-1 and Grok 2 remain available as open weights.
Further deprecations and shutdowns: grok-voice-transcribe-1.0 reached end of life on 2 October 2026, and requests run on grok-voice-transcribe-2.0. On 14 October 2026, xAI deprecated grok-voice-2.0; since then the alias grok-voice-latest points to grok-voice-3.0. On 2 November 2026, grok-imagine-image-quality will be shut down and replaced by grok-imagine-image-2.0 with quality low.
On Amazon Bedrock, the Grok models each have a runtime of at least one year from launch (Grok 4.3 until at least 15 June 2027, Grok 4.6 until 18 August 2027, Grok 4.7 until 28 September 2027). xAI has not announced a Grok 5 in its documentation and release notes so far.
Integration with CompanyGPT
Our CompanyGPT solution offers a GDPR-compliant alternative that combines different models while ensuring the highest data protection standards. For companies that want to stay on the safe side, this is often the better choice.
Our recommendation
Grok 4.7 is the default for new Grok projects via the xAI API: same price as Grok 4.6, newer base model and a knowledge cutoff of May 2026. For cost-sensitive mass processing with long documents, Grok 4.3 with 1 million tokens of context and a batch discount is the cheaper choice, and for pure coding Grok Build 0.1. Grok 4.7 is GA on Microsoft Foundry and Google Cloud, and usable on Amazon Bedrock via the US Geo or Global profile. Model IDs should be configured centrally, because only a few weeks lay between Grok 4.3, 4.5, 4.6 and 4.7.
Grok is only suitable to a limited extent for personal data. Neither the xAI API nor Microsoft Foundry, Amazon Bedrock or Google Cloud offers documented processing exclusively in the EU for Grok. If you use Grok anyway, you should enable Zero Data Retention, not transmit personal data or carry out a data protection impact assessment. For applications with EU data residency, these alternatives are more suitable:
| Application | Alternative |
|---|---|
| General LLM tasks with EU hosting | Anthropic Claude or OpenAI GPT |
| European provider | Mistral |
| Open source and self-hosting | Meta Llama or Qwen |
| GDPR-compliant complete solution | CompanyGPT |
For individual advice on the right AI strategy and GDPR-compliant implementation, contact innFactory AI Consulting.