xAI Grok

innFactory AI Consulting from Rosenheim advises companies in the DACH region (Germany, Austria, Switzerland) on GDPR-compliant integration of Grok. With the new flagship Grok 4.6 (12 August 2026), its predecessor Grok 4.5 (8 July 2026), the General Availability of Grok 4.3 (May 2026) and the coding model Grok Build 0.1, the vendor has significantly expanded its portfolio in 2026 – and restructured itself at the same time: xAI has been merged with SpaceX and, per its own statements and media reports, now operates as SpaceXAI; the coding company Cursor was acquired. This overview is current as of 22 August 2026.

New in August 2026: Grok 4.6

On 12 August 2026, SpaceXAI introduced Grok 4.6 – the current frontier model for coding, agentic tasks and knowledge work. Per the vendor, the emphasis is on long-running agents and multi-step, interactive work: from researching and analysing information to working across a whole codebase and turning a product idea into a functioning application.

Key facts per the xAI documentation:

• Context window: 500,000 tokens, text and image input, text output with no documented output limit
• Reasoning effort: four levels – low, medium, high (default) and xhigh
• API pricing: $2 input / $0.50 cache read / $6 output per 1M tokens for prompts below 200k tokens; from 200k tokens $4 / $1 / $12
• Knowledge cutoff: 1 February 2026
• Benchmarks (vendor-provided): per xAI, 61 points on the Artificial Analysis Intelligence Index, plus CursorBench 3.2 69.9%, DeepSWE 1.1 65.9% and FrontierCode 1.1 61.3%

Grok 4.6 is available through the SpaceXAI API and console, in Cursor and Grok Build, and via partners including OpenRouter, Vercel and Cloudflare. Also new in August: Grok Bot – per the release notes, durable agents on persistent cloud infrastructure with messaging, approvals and connectors – is generally available.

Grok 4.6 on Amazon Bedrock – with an important caveat

Since 19 August 2026, Grok 4.6 is generally available on Amazon Bedrock (model launch date per the model card: 18 August 2026). That is a meaningful opening, because until then Grok was effectively unusable for Bedrock customers. For the GDPR assessment, however, how Bedrock offers the model is decisive:

• There are two cross-Region inference profiles: us.xai.grok-4.6 (US Geo, routing exclusively within the United States) and global.xai.grok-4.6 (worldwide routing).
• Per AWS documentation, an EU geo profile does not exist. In the EU Regions (including Frankfurt, Ireland, Stockholm, Zurich, Paris, Milan, Spain and London) only the Global profile is enabled.
• In-Region inference via bedrock-runtime is not supported; only via the bedrock-mantle endpoint in us-west-2.

What this means for EU companies: you can call Grok 4.6 from a European AWS account – but processing is not confined to the EU. Anyone who needs EU data residency cannot achieve it for Grok via Bedrock today.

Bedrock pricing (Standard tier, per 1M tokens): Global CRIS $2.00 input / $6.00 output / $0.50 cache read; US Geo $2.20 / $6.60 / $0.55. The Priority (2x) and Flex (0.5x) tiers are supported; a Reserved tier is not.

Grok 4.5: the previous flagship (July 2026)

On 8 July 2026, SpaceXAI released Grok 4.5 – following a private beta at SpaceX and Tesla since late June. The model is built on the V9 foundation architecture with roughly 1.5 trillion parameters and was positioned as a “workhorse” for coding, agentic tasks and knowledge work; Elon Musk described it as an “Opus-class model, but faster, more token-efficient and lower cost”.

The key facts:

• Context window: 500,000 tokens
• API price: $2 input / $0.30 cache read / $6 output per 1M tokens below 200k tokens; above that $4 / $0.60 / $12
• Features: reasoning levels (low/medium/high), prompt caching, context compaction for long agent loops, function calling, web/X search, code execution
• Benchmarks (vendor-provided): among others 83.3% on Terminal-Bench 2.1 and 62.0% on DeepSWE 1.0; an independent Snorkel evaluation (GDPVal+) put Grok 4.5 at a 29% pass rate ahead of GPT-5.5 (22%) and Opus 4.8 (21%). Classic benchmarks such as MMLU or GPQA were not published at launch.
• Token efficiency: per the vendor, ~4x fewer output tokens per SWE-Bench Pro task than Opus 4.8 (max)

EU availability update: at launch, per the official docs, Grok 4.5 was not available in the API console for EU users. Per the xAI release notes, access for EU users was enabled during July 2026. Grok 4.5 is still not listed on Microsoft Foundry.

New in May/June 2026

Grok 4.3 is now Generally Available

Following the beta launch in April 2026, xAI moved Grok 4.3 into general rollout in early May 2026. In addition to multimodal video understanding and the 1M token context window, the GA version brings:

• Custom Skills: persistent expertise (formatting rules, workflow steps, document styles) that Grok automatically applies across every conversation
• Integrated code execution environment: Grok can write code, run it, install dependencies, and produce real files

Grok Build 0.1: Dedicated Coding Model

On 29 May 2026, xAI released Grok Build 0.1 in public beta on the xAI API. The model is specifically trained for agentic coding:

• 256K token context window
• Text and image input (e.g. UI mockups, architecture diagrams, error screenshots)
• Native MCP support and integration with Cursor, Kilo Code, OpenCode and others
• Inference speed > 100 tokens/second
• Pricing: $1 per 1M input tokens, $2 per 1M output tokens, cache read $0.20 per 1M tokens

Technical Strengths of Grok 4.3

Extended Context Processing

With a context window of 1 million tokens, Grok 4.3 ranks among the most powerful models for processing extensive documents. For enterprises, this means:

• Complete analysis of entire codebases without splitting
• Processing comprehensive contracts and technical documentation in one pass
• Consistent analysis of long conversation histories and protocols

Benchmark Results

Current performance tests show strong results:

• MMLU (Multitask Understanding): 89.5% - on par with leading models
• LiveCodeBench: 79.4% with tool use - surpasses many established competitors
• SWE-Bench (Software Engineering): 75% - leading in real-world coding tasks
• GPQA (Graduate Science): 88% - outstanding in scientific questions

Agentic Capabilities

Grok 4.3 offers extended capabilities for autonomous multi-step tasks:

• Reduced hallucination rate by 65% compared to predecessors
• Improved generalization of programming logic across language boundaries
• Native integration of web and X search for current information

EU Availability and GDPR Compliance

Microsoft Foundry (formerly Azure AI Foundry)

Per Microsoft documentation (as of 20 August 2026), the Foundry catalogue offers the following Grok models “sold by Azure”: grok-4.3 (preview), grok-4-20-reasoning and grok-4-20-non-reasoning (preview), grok-4.1-fast-reasoning and -non-reasoning, grok-4 and grok-code-fast-1. Registration is required for grok-4 and grok-code-fast-1.

The current models Grok 4.5 and Grok 4.6 are not listed there. Anyone wanting to run Grok under Microsoft governance is therefore working with an older generation. Regional availability differs by model and deployment category and should be checked in the Microsoft documentation before deciding.

AWS Bedrock: available, but without EU geo inference

Grok 4.3 has been listed on Amazon Bedrock since June 2026, and Grok 4.6 since 19 August 2026. For GDPR purposes the picture is nuanced:

For comparison: for Anthropic models, Bedrock offers an EU geo inference profile that routes requests exclusively within the EU geography. No such profile exists for Grok so far.

Google Cloud

Grok models remain unavailable on the Gemini Enterprise Agent Platform (formerly Vertex AI) – including Frankfurt (europe-west3).

Practical conclusion: there is currently no clean route to running Grok with strict EU data residency. Microsoft Foundry carries only older generations, Bedrock offers no EU routing, and Google Cloud carries no Grok models at all. For GDPR-sensitive processing we therefore continue to recommend the alternatives in the next section.

Pricing

Prices per the SpaceXAI model documentation (as of 22 August 2026), given as input / cache read / output per 1M tokens. From a prompt length of 200,000 tokens, every model is billed at double the rate.

Additional modalities per the price list: Grok Imagine Image 2.0 ($0.04 per image), Grok Imagine Video 1.5 ($0.080 per second) and the Grok Voice API ($0.08 per audio minute).

On Amazon Bedrock (Standard tier, per 1M tokens): Grok 4.6 costs $2.00 input / $6.00 output / $0.50 cache read via the Global profile, and $2.20 / $6.60 / $0.55 via the US Geo profile. The Priority tier is billed at 2x and the Flex tier at 0.5x.

Tool Calls:

• Web/X search, code execution: $5 per 1,000 calls
• Batch API: 50% discount for asynchronous processing

Use Cases for DACH Enterprises

Technical Analysis

• Software engineering tasks with complete codebase understanding
• Automated code reviews and refactoring suggestions
• Technical documentation analysis

Scientific Applications

• Processing extensive research documents
• STEM-related questions and calculations
• Graduate-level scientific analysis

Social Media and Trend Monitoring

• Integration with X/Twitter for real-time data analysis
• Content creation with current context
• Trend identification and market observation

Critical Assessment

Ethical and Practical Concerns

Despite technical strengths, concerns about Grok persist:

• Controversies: Connection to Elon Musk and political positions
• Minimal security filters: Can be problematic for regulated industries
• No EU data residency: Microsoft Foundry carries only older generations, Bedrock offers no EU geo profile for Grok, and Google Cloud carries no Grok models at all
• Knowledge cutoff: Trained until 1 February 2026 (Grok 4.6), current only via web search

Better Alternatives for Enterprise Use

For professional applications in the DACH region, we often recommend:

Integration with CompanyGPT

Our CompanyGPT solution offers a GDPR-compliant alternative that combines various models while ensuring the highest data protection standards. For companies that want to stay on the safe side, this is often the better choice.

Outlook

The foundation model announced in spring as “Grok V9-Medium” (1.5 trillion parameters) shipped as Grok 4.5, and its successor Grok 4.6 followed just over five weeks later. As of 22 August 2026, Grok 5 has not been released; per media reports the model is still in training and no confirmed date exists. Reports of a larger Grok 4.7 in preparation also come from secondary sources. We treat both as unconfirmed.

The cadence is notable: only a few weeks separated Grok 4.3 (May), 4.5 (July) and 4.6 (August). For production integrations that means configuring model IDs centrally rather than scattering them through application code – otherwise every generation change turns into a release.

Our Recommendation

Grok 4.6 has been the new default in the Grok portfolio for coding and agent tasks since 12 August 2026 – with a 500k context window, four reasoning effort levels and attractive pricing. Since 19 August 2026 it is also available on Amazon Bedrock, which simplifies governance and billing for AWS customers. Grok 4.3 remains the proven all-rounder with its 1M token context window, Custom Skills and code execution; for pure coding workloads, the cheaper Grok Build 0.1 is an alternative.

However: The ethical concerns, the minimal security filters and above all the absence of EU data residency make Grok a risky choice for many companies in regulated environments. Bedrock availability does not change that as long as there is no EU geo inference profile. For business-critical processing and personal data, we recommend established alternatives with stricter governance standards and demonstrable EU data handling.

For individual consulting on the appropriate AI strategy and GDPR-compliant implementation, contact innFactory AI Consulting.