On 7 August 2026, SpaceXAI — the company formerly known as xAI — released Grok 4.6. Within a week it was sitting near the top of the independent benchmark tables, at roughly a third of the price of its closest competitors.
That combination gets attention. It also raises a question we keep getting from Belgian clients: can we actually use this?
The short answer is that Grok 4.6 is a genuinely strong model with a genuinely unresolved regulatory position in Europe. Those are two separate facts, and it is worth keeping them separate.
What actually shipped
Grok 4.6 is a 1.5-trillion-parameter model built on the same V9 foundation as Grok 4.5. That detail matters more than it sounds: this was not a scale-up. SpaceXAI kept the base model and pushed the gains through improved supervised fine-tuning and reinforcement learning instead.
The practical specifications:
| Grok 4.6 | |
|---|---|
| Context window | 500,000 tokens |
| Input | Text and images |
| Output | Text only |
| API price | $2 per million input tokens, $6 per million output |
| Cached input | $0.50 per million (prompts under 200K) |
| Long-prompt penalty | Prompts over 200K tokens bill the entire request at $4 / $12 |
That last row is easy to miss and expensive to learn. Cross the 200K threshold and you do not pay the higher rate on the overage — you pay it on everything, input and output. If your workload sits near that line, it is cheaper to stay deliberately under it than to drift over.
SpaceXAI also released grok-voice-think-fast-2.0 on 5 August, a speech-to-speech model that grok-voice-latest now routes to. And a larger 2.1-trillion-parameter Grok 4.7 is expected within weeks — worth knowing before you invest heavily in tuning against 4.6 specifically.
Where it lands against the competition
Artificial Analysis, which runs independent evaluations, measures Grok 4.6 at 61 on its Intelligence Index. That puts it level with GPT-5.6 Sol, about five points above Grok 4.5, and just behind Anthropic’s Claude Opus 5 and Claude Fable 5.
On AA-Briefcase — a private benchmark specifically for long-horizon agentic knowledge work, which is much closer to what most business use looks like than a reasoning quiz — Grok 4.6 debuts with an Elo of 1577. Strong, and still behind the Claude Opus 5 family.
The honest summary: Grok 4.6 is at the frontier without leading it, and it is priced well below the models it matches. For raw cost-per-capability, that is a real achievement.
The part that should stop a Belgian company
Grok is currently the only widely-used AI assistant under simultaneous investigation by multiple EU supervisory authorities over the lawfulness of its foundational training data.
The specifics:
- Ireland’s Data Protection Commission, France’s CNIL, and the UK’s ICO are all investigating xAI/SpaceXAI’s use of X user data to train Grok without an established lawful basis. Between late 2023 and early 2026, the company’s training strategy relied on large-scale scraping of public X posts.
- In September 2024, X agreed to permanently stop using EU and EEA users’ public posts from May–August 2024 for Grok training, after suspending that use in August 2024. That settles one window. It does not settle the rest.
- In January 2026, users discovered Grok would generate non-consensual sexualised images of real people from ordinary photos. Ireland’s DPC opened a large-scale inquiry into X over those images in February 2026.
- The European Commission has opened a formal Digital Services Act investigation into X’s management of systemic risks connected to Grok.
None of these have concluded. That is precisely the problem: an unresolved investigation is not an acquittal, and you cannot plan around a finding that has not landed.
There is also a quieter operational issue. EU account access and EU data residency are two different things. Grok 4.5 opened to EU users on the API console in July 2026, while its model page still listed regions as us-east-1 and us-west-2 only. Being able to log in from Brussels does not mean your data stays in Europe.
So what can you actually do with it?
We would not put personal data of EU data subjects through Grok right now. Not because a regulator has ruled against it, but because several are actively asking the question, and “we were waiting for the outcome” is a poor answer to give a client whose data was involved.
That still leaves real ground:
Reasonable today
- Evaluating the model against your own tasks, using synthetic or non-personal test data
- Code work that does not include customer data, credentials, or production records
- General research, drafting, and summarisation of public material
- Benchmarking — genuinely useful to know how it compares on your workload, not a leaderboard’s
Not reasonable today
- Anything involving customer records, HR data, patient data, or personal data of EU data subjects
- Workflows in a regulated context (finance, healthcare, public sector) without explicit legal sign-off
- Anything where you would struggle to explain the data-processing chain to a client
If you do decide to deploy it beyond evaluation, the minimum homework is unchanged from any other processor: a signed data processing agreement under Article 28 GDPR, a clear answer on international transfers under Chapter V, the current sub-processor list, and documented data residency. Ask for these in writing before, not after.
For teams that want the enterprise controls, Grok Business runs at $30 per user per month, and Grok Enterprise is negotiated — it adds an “Enterprise Vault” with dedicated infrastructure, application-level encryption, and customer-managed keys. Those are meaningful controls. They address confidentiality; they do not address the training-data question, which sits upstream of anything your contract can fix.
One more thing that changed this month
Since 2 August 2026, the EU AI Act’s Article 50 transparency obligations apply. If you deploy a chatbot that interacts with people, or you publish AI-generated content, you now have disclosure and labelling duties — regardless of which model sits behind it. That is a separate obligation from GDPR and it applies to Grok, GPT, Claude and everything else equally.
We wrote that up separately, because it catches more companies off guard than the model choice does.
Our take
Grok 4.6 is the best price-to-capability ratio at the frontier right now, and SpaceXAI is shipping fast enough that 4.7 will likely reset the comparison again before autumn.
For a Belgian SME, though, model quality is rarely the binding constraint. The binding constraint is whether you can explain your data chain to a client, an auditor, or a DPA. On that measure Grok is, for now, the hardest of the major options to defend — and the gap between it and the alternatives on raw capability is small enough that the trade is not obviously worth it.
Evaluate it. Benchmark it. Keep personal data out of it until the investigations resolve.
What’s next?
- AI tools for Belgian SMEs — a practical overview for 2026 — how the major platforms compare on price, features and GDPR posture
- Claude vs ChatGPT vs Gemini: document creation compared — a hands-on comparison for real document work
Wondering how this applies to your own setup? Get in touch — we help Belgian organisations pick and deploy AI tooling that survives a compliance review.