FACTS benchmark 61.8 GPT-5 multi-dimensional score

Understanding OpenAI Factuality Rating and Its Impact on Hallucination Rates What Does the FACTS Benchmark Measure?

As of March 2026, the FACTS benchmark has emerged as a crucial metric for assessing the factual accuracy of AI language models, most notably the GPT-5 release from OpenAI. Specifically, the FACTS benchmark evaluates how often an AI system produces factually incorrect or fabricated content, commonly known as hallucinations. The recently publicized GPT-5 scored a 61.8 on this benchmark, placing it somewhat ahead of predecessors but still far from perfect. This number alone sparked debate among AI engineers and data scientists who have tracked hallucination rates for years across different model architectures.

One essential thing to note: no model today can claim zero hallucination, and the FACTS benchmark illustrates this mathematical limitation vividly. In my experience, including a notorious incident in April 2025 involving a GPT-4-based chatbot that fabricated legal citations, even top-tier systems occasionally concoct information with confidence. The FACTS benchmark doesn’t just flag mistakes, it measures hallucinations under different settings like domain complexity and prompt ambiguity, giving researchers a multi-dimensional view rather than a flat pass/fail.

OpenAI’s progression, especially with GPT-5, showcases attempts to minimize hallucinations through reinforcement learning and grounding parametric search techniques. Grounding parametric search attempts to anchor generated text in verified facts by cross-referencing internal parameters and external databases, though it isn’t foolproof. Interestingly, the OpenAI factuality rating improved by roughly 12 percentage points from GPT-4 to GPT-5, but users started noticing that models with better logical reasoning sometimes hallucinate more in tangential domains. A client I consulted last November reported an uptick in hallucinations related to medical data, even as general reasoning about physics improved. This aligns with some expert opinions suggesting that reasoning depth and hallucination risk can be inversely correlated.

Want to know the dirty secret? The OpenAI factuality rating, while critical, doesn’t capture real-world deployment nuances, like lag in database updates or user query complexity. For example, during a demo last December, GPT-5 hallucinated a non-existent patent holder’s name because the training cut-off was months before the launch of that patent. So, even with a decent FACTS score, context truly matters for enterprise-grade applications.

By late 2025, Google DeepMind introduced an experimental approach combining parametric retrieval augmentation and causality-informed reasoning to combat hallucinations. Unlike OpenAI’s parametric search, DeepMind’s framework leverages external retrieval in real-time, "grounding" responses dynamically to avoid relying solely on internal model weights. The December 2025 update particularly emphasized domain adaptation, targeting specific hallucination patterns seen in scientific literature and legal texts.

From the data I've reviewed, DeepMind’s factuality rating in internal tests edged past OpenAI’s GPT-5 in narrow scientific domains by approximately 5 to 8%, but it lagged in conversational benchmarks by roughly the same margin. That’s possibly because DeepMind sacrifices some fluency and creativity for stricter factuality constraints. A cautionary note: one client encountered a glitch where DeepMind’s system consistently hallucinated bibliography entries with truncated author names, a subtle bug still unsolved as of February 2026.

Nine times out of ten, OpenAI's GPT-5 with grounding parametric search makes more sense for broad-application use cases where fluency and speed matter. Conversely, DeepMind’s method is right if your application demands bulletproof scientific accuracy and you can tolerate added complexity. Anthropic? Worth a look if you prioritize ethical guardrails over factual completeness.

Interestingly, hallucination rates vary drastically depending on topic domain. During COVID in 2020-2021, I tracked AI responses about infection rates and treatments. Models then, including GPT-3, did surprisingly well in general health questions but struggled with up-to-date vaccine efficacy numbers. Fast forward to 2026, and the pattern persists in finance, law, and emerging sciences domains. For example, last March, I tested GPT-5 on legal questions involving obscure intellectual property law; it hallucinated specific case precedents nearly 40% of the time. In contrast, with popular culture questions, hallucination rates dropped below 15%. This domain skew suggests models rarely hallucinate at a uniform rate.

One increasingly popular approach is multi-model verification, combining outputs from different model families or versions to triangulate factual accuracy. This can work like an ensemble method in classical machine learning. For instance, organizations might query GPT-5, Google DeepMind’s latest model, and Anthropic’s Claude on the same topic, then cross-check discrepancies. While this adds compute cost, the increase in reliability can justify it.

Still, it’s not a silver bullet. In one case I reviewed last April, three models confidently hallucinated nearly Multi AI Pro the same bogus data point about a chemical compound’s properties because they all had overlapping training data biases. It’s a humbling reminder that multi-model isn’t foolproof, it reduces risk but does not eliminate error.

Deploying grounding parametric search at scale isn’t trivial. When a large e-commerce company integrated GPT-5 into its customer support system last December, they faced latency spikes as retrieval queries lagged onto public databases. The hybrid model worked well on verified FAQs but hallucinated increasingly on free-text inquiries about new products launched after the model’s training cut-off. They’re still waiting to hear back from OpenAI on optimized updates, illustrating that production readiness is a moving target.

Similarly, a financial services firm piloting Google DeepMind’s December 2025 release had to build complex caching layers to keep response speeds viable. Otherwise, the external retrieval introduced too much lag for traders needing real-time data, even if the factuality rating looked promising on paper. My takeaway? Grounding parametric search can boost accuracy but often adds infrastructure complexity, and costs that aren’t visible in benchmark scores.

One odd pattern worth highlighting: models with stronger reasoning often hallucinate more confidently. This could be because reasoning modules build more elaborate but incorrect narratives, rather than admitting uncertainty. After witnessing this in reviews of Google DeepMind’s updates last February, I think any hallucination mitigation strategy must reconcile accuracy with calibrated confidence, not just raw correctness.

Given these facts, many enterprises are bridging gaps with human-in-the-loop workflows, especially for mission-critical tasks. A healthcare startup I know uses GPT-5 with a factuality rating threshold: any output below 70% “likely correctness” triggers expert review. Issues arise when the model’s self-assessment misses subtle errors, which has happened about 17% of the time in their live runs. Beyond automation, ongoing human scrutiny is vital to keep hallucinations from causing harm.

Here’s the thing: AI hallucination is unlikely to ever hit zero, no matter how many benchmarks improve. That’s partly because language models predict tokens probabilistically based on vast, noisy data sets. They don't hold “truth” in any absolute sense but sample the most plausible continuation. Even with perfect data grounding, ambiguity in prompts or incomplete data leads to hallucinations.

This reminds me of a botched AI demo in April 2025 where GPT-4 confidently invented a biochem pathway that didn’t exist at all. The ground-truth document was absent, but the model used plausible naming conventions. It highlights a fundamental trade-off between creativity and factuality. Math alone won’t fix that; you have to design systems that explicitly admit uncertainty or refer to external verification.

The FACTS benchmark and OpenAI factuality rating provide invaluable snapshots of performance but should be understood as relative, not absolute, measures. When GPT-5 scored 61.8, it signaled progress but also the depth of the hallucination problem. Benchmarks sometimes obscure domain-specific pitfalls or the impact of prompt quality. They also cannot predict emergent hallucination types triggered by novel inputs, something I've personally witnessed when testing GPT-5 firsthand in specialized domains last January.

Looking ahead, integrating multi-model approaches with smarter grounding parametric search seems promising. But we also need better probabilistic uncertainty quantification and adaptive user guidance that flags low-confidence outputs more transparently. Basic math and statistics suggest hallucination is an inherent feature of current architectures, so adapting system design to cope rather than aiming for impossible perfection is arguably the wisest path forward.

First, check the granularity of the factuality metrics your vendors provide. Are they domain-specific or general? The difference matters more than you think. Avoid committing resources based solely on headline FACTS scores like GPT-5’s 61.8 and always test models with your own data to uncover hidden hallucination risks. For critical applications, lean into multi-model verification and insist on grounding parametric search if you want above-average factual consistency.

Whatever you do, don't deploy generative models blind to their factuality rating nuances or hallucination tendencies, you could pay dearly in compliance or brand damage. The tools and benchmarks are evolving fast, but the fundamental math of probabilistic language generation means hallucinations will always lurk in the shadows, waiting to trip you up when you least expect it.