More Facts, More Errors: Why a Better AI Model Isn't Automatically More Trustworthy

On July 8, 2026, xAI – referred to in parts of the press as "SpaceXAI" – launched its new model, Grok 4.5, worldwide. The headlines were quickly written: rank 4 in the Artificial Analysis Intelligence Index, a jump of 16 points over its predecessor Grok 4.3, trailing only Claude Fable 5, GPT-5.5, and Claude Opus 4.8. An impressive leap forward, you might think, if you're currently in the market for a new AI model for your business.

But if you look more closely at Artificial Analysis's raw data, you find a second story there that doesn't make the front screen of most launch articles: Grok 4.5's hallucination rate more than doubled on a separate benchmark specifically designed to measure factual accuracy. That contradiction is exactly why we at Amazing Agents are looking closely at the Grok 4.5 case – not because we think the model is particularly good or bad, but because it illustrates, textbook-style, how sharply you need to distinguish between "intelligent" and "trustworthy" AI before you deploy it productively in your business.

For you as an owner or manager in the Mittelstand, this is more than a technical footnote. If you're considering running a customer service chat on a new model, having offers automatically drafted, or letting AI produce first drafts for contracts and legal questions, this exact distinction between intelligence and trustworthiness decides whether you save time with it or take on new liability risks. The Grok 4.5 case makes a good teaching example because, in just a few weeks, it made several aspects of this debate visible at once: benchmark ranking, factual accuracy, regulation, and price.

The Grok 4.5 case: more knowledge, more confidence, more errors

Technically, Grok 4.5 is a significantly larger model than its predecessor: 1.5 trillion parameters, roughly three times as many as Grok 4.3, according to Artificial Analysis. At the same time, its context window has shrunk from 1 million to 500,000 tokens – a detail that hardly appears in any launch coverage, but could matter to you if you want to work with long contract documents, extensive minutes, or entire case files.

On Artificial Analysis's general Intelligence Index, which reflects a broad mix of reasoning, coding, and knowledge tasks, Grok 4.5 scores 54 points and lands at rank 4 on the current leaderboard. That's solid progress over Grok 4.3. But the Intelligence Index doesn't measure how reliable a model is on factual questions. For that, Artificial Analysis has developed a dedicated, specialized benchmark: the AA-Omniscience Index.

And that's exactly where the truly remarkable pattern of this model update shows up: accuracy on factual questions rises from 35 percent (Grok 4.3) to 52 percent (Grok 4.5) – genuine progress. But at the same time, the hallucination rate rises from 25 percent to 54 percent. So the model knows more correct things – and, at the same time, is also wrong noticeably more often, with great confidence. The two together don't cancel out to zero: a model can become both more competent and less reliable at once, if it learns to give an answer more often instead of staying silent when uncertain.

What the AA-Omniscience Index actually measures

To properly assess this effect, it's worth looking at the methodology. The AA-Omniscience Index is deliberately built to reward correct answers, penalize confidently wrong answers – and explicitly not penalize a refusal to answer. A model that honestly says "I'm not sure about this" on an uncertain question therefore scores better on this benchmark than a model that states a wrong number, a wrong citation, or a wrong legal provision with full confidence.

This exact design reveals what pure intelligence rankings obscure: a model can "know more" while simultaneously being wrong with more confidence.

Imagine a customer asks your chat about a delivery deadline, a warranty condition, or a legal question the model has no reliable information about. A model that honestly signals "I can't answer that with certainty, I'll get back to you" costs you, at worst, some waiting time. A model that instead confidently states a wrong delivery deadline or an invented warranty condition can cost you a customer, a complaint, or a legal dispute in the worst case. For your practice in the business, that means: a chatbot that would rather stay silent or ask a follow-up question when uncertain is often more valuable in a customer-facing context than one that always delivers a fluently worded, convincing-sounding answer – even if that answer was wrong in more than every second case in the AA-Omniscience test. Important to note: this 54 percent hallucination rate comes from a very specific benchmark focused on factual knowledge. Industry-wide figures for hallucination rates in open-ended fact generation without retrieval tend to sit between 3.1 and 19.1 percent for frontier models in 2026, between 17 and 88 percent for domain-specific tasks like legal research, and up to 64 percent for medical case summaries without countermeasures. These figures aren't directly comparable to the AA-Omniscience value, since different methodologies underlie them – but they show that Grok 4.5, with its result, sits at the upper end of the risk spectrum even by industry-wide comparison.

The ranking fallacy: why rank 4 says nothing about trustworthiness

The real thinking error we observe in many of our conversations with Mittelstand businesses isn't in the specific assessment of Grok 4.5, but in a much more general assumption: that a better ranking automatically means a more trustworthy model. The Grok 4.5 example refutes that assumption quite plainly. A model can rise in the general Intelligence Index while performing noticeably worse on a specific trust benchmark at the same time.

So if you're selecting an AI model for your business – whether for the customer service chat on your website, for automated quote generation, or for first drafts in legal advice – a glance at the position in a general ranking isn't enough. What matters is the type of error: does the model refuse to answer when uncertain, or does it confidently invent facts? The Intelligence Index doesn't make that distinction, but a specialized fact benchmark like AA-Omniscience does.

Grok 4.5's agentic capabilities also show a mixed picture that confirms this nuance. On the GDPval-AA-v2 benchmark, the model achieves an Elo score of 1543, landing between Claude Opus 4.8 (1600) and GLM-5.2 (1513). On the tau3-Banking benchmark, Grok 4.5 scores 33 percent versus 31 percent for GPT-5.5 – a narrow lead. On the Coding Agent Index, Grok 4.5 ties with GPT-5.5 and Codex at a score of 76, but trails Claude Fable 5. These figures show that Grok 4.5 is quite competitive on agentic coding and automation tasks. That's exactly the point: for internal, technical workflows the model can be a good choice – for customer-facing or legally sensitive applications, the high hallucination rate is a serious warning sign.

The EU delay: regulation as a second layer of control

Another detail surrounding the Grok 4.5 launch deserves your attention, especially if you run a company in Germany or another EU member state: xAI did not enable Grok 4.5 in the EU at launch. The model is blocked in all 27 EU member states. Several sources cite Grok 4.5's classification as a general-purpose AI model with systemic risk under the EU AI Act (Regulation 2024/1689) as the reason. That classification requires mandatory model evaluations, adversarial testing, incident-reporting processes, and cybersecurity assessments before an EU deployment. An EU launch in mid-July 2026 was floated, but reports treat it more as a target date than a firm commitment.

For you as a Mittelstand business owner, this is more than a footnote. It illustrates that regulatory requirements represent an additional hurdle in model selection for European companies regardless of how well a model performs on an international leaderboard. If a provider is only allowed to release a model in the EU after adversarial testing and cybersecurity assessments, that is, at its core, the same logic that should apply inside your own business: before an AI system operates independently with customers, contracts, or sensitive data, it needs a control layer that checks more than raw capability alone.

The cost trade-off: cheap, and precisely for that reason dangerous

One aspect that makes this article's core thesis especially clear is Grok 4.5's price. According to Artificial Analysis, the model costs $2 per million input tokens and $6 per million output tokens, with a cache discount down to $0.50. Translated into concrete tasks, costs run around $0.31 per task on the Intelligence Index benchmark, and around $2.59 per task on the Coding Agent Index. That makes Grok 4.5 significantly cheaper than Claude Fable 5 or GPT-5.5.

This price difference creates exactly the conflict this article is about: a model that is attractive on price for many Mittelstand businesses, but that can become especially dangerous for that very reason when deployed without a control layer. Low per-request costs tempt you to skip validation steps – after all, "each request barely costs anything." But that logic overlooks that a 54 percent error rate on the fact benchmark suggests exactly the opposite: the cheaper and more frequently a model is used in customer-facing or legally sensitive processes, the more important a structural review layer becomes.

Several industry articles on Grok 4.5 now phrase this as a clear recommendation: for "legal, financial or client-facing writing," every claim that can't be independently verified should be checked by a human. The recommendation specifically is to use Grok 4.5 for internal coding and agentic workflows, but not for customer-facing or regulated applications without an additional control layer. That's fundamentally nothing other than the human-in-the-loop logic that a properly built agent – unlike raw chat access to a language model – can structurally enforce: through approval steps, citation requirements, clearly defined escalation rules.

What this means for your Mittelstand business

The obvious conclusion isn't to avoid new, capable models like Grok 4.5. The conclusion is to never let them near your customers, your contracts, or your bookkeeping unfiltered. In practice, a tiered model has emerged for this, discussed in several industry sources for 2026: from L0 – pure observation, the model suggests, a human decides everything – up to L4, full autonomy without human oversight. Most Mittelstand workloads are realistically classified at L2 (preparation with approval) or L3 (execution with a defined veto window), not L4.

Concretely, that means: an agent that prepares your quotes, checks invoices, or pre-sorts customer inquiries can and should use a capable, even cost-effective model in the background – as long as the architecture built around that model ensures that uncertain or consequential statements go to a human for approval before they go out or move money.

This control layer also matters because governance reality in many companies lags behind the technological development. According to industry research, only around one in five companies currently has a truly mature governance model for deploying autonomous AI agents – even though far more companies want to introduce and adapt such agents. That gap between ambition and control is exactly where incidents like Grok 4.5's doubled hallucination rate become dangerous in practice: not because the model is bad, but because it's deployed without the right framework around it.

If you're introducing AI agents in your company or want to replace existing chat access with structured workflows, it's therefore always worth asking about the architecture before the model choice: where in the process do you need a citation requirement? Where does an approval step belong? Where does an escalation to a human need to happen before any answer even leaves the company? These questions are independent of whether you end up deploying Grok 4.5, Claude Fable 5, GPT-5.5, or another model – but they decide, to a large degree, whether a more capable model becomes a productivity gain for you or a risk.

A concrete practical example illustrates the principle: use a model like Grok 4.5 for internal research and preparing a quote, and the risk is manageable, because an employee reviews the result anyway before it goes to a customer. Let the same model communicate directly with a customer without any intermediate step, or draft an automated legal opinion, and the risk shifts immediately onto your company and your customers. The difference doesn't lie in the model, but exclusively in the architecture you build around it – and that's exactly the core of a well-designed agent, as opposed to open chat access.

In the end, the Grok 4.5 case is less a story about a single model than about a habit of thought you should break when choosing AI: taking the ranking as the sole yardstick. A higher Intelligence Index score tells you how smart a model performs on a broad mix of tasks. It tells you nothing about whether it stays silent when uncertain or delivers the wrong answer with full confidence. You need to know exactly that difference before deploying a model in a customer-facing or regulated process – and that's exactly why you need an architecture that doesn't leave control to chance.