LLM hallucination statistics 2026: Model comparisons and benchmarks
Key takeaways:
- Modern LLMs still hallucinate between 4% and 80%+ of the time depending on the task and benchmark.
- Google DeepMind's Gemini 4 Argon leads our blended ranking at 15.43%, followed by Grok 4.7 (25.71%) and GPT 6 Astra (27%).
- Web search and RAG cut hallucination rates by a median of 30% on HalluHard, though they don't eliminate ungrounded claims entirely.
- Newer model versions usually improve over their predecessors, but occasional regressions (such as Claude Opus 4.6 on FACTS v2) mean every upgrade still needs testing.
Do modern LLMs still hallucinate?
Modern Large Language Models (LLMs) still hallucinate at a rate of 4 to more than 80 percent depending on the use case and method of measurement. Hallucinations occur when directly prompting the model, in summarization, and even when the model is grounded through search. This means that RAG is not an effective safeguard against hallucinations.
Our statistics hub shows hallucination rates across the latest models, different benchmarks, and in specialized domains like legal and life sciences.
Which LLM has the lowest hallucination rate?
There is no universal rate that would identify a single model as having the lowest hallucination rate. Different benchmarks come to different conclusions.
Lowest hallucination rates: the top 10
Median blend across available benchmarks · lower is better
- xAI
- OpenAI
- Anthropic
- Z.AI
Swipe or scroll horizontally to see the full chart.
| Rank | Model | HalluHard + search | AA v1.1 | FACTS failure | Median blend |
|---|---|---|---|---|---|
| 1 | Gemini 4 Argon | — | 4.26% | 26.60% | 15.43% |
| 2 | Grok 4.7 | — | 20.01% | 31.40% | 25.71% |
| 3 | GPT-6 Astra | 27.23% | 49.78% | 19.90% | 27.23% |
| 4 | Gemini 3.5 Flash-Lite | — | 26.71% | 28.80% | 27.76% |
| 5 | GPT-5.4 | 31.88% | — | 29.70% | 30.79% |
| 6 | Claude Opus 4.5 | 30.20% | — | 37.90% | 34.05% |
| 7 | GPT-6.1 Sol | — | 51.88% | 22.00% | 36.94% |
| 8 | GPT-6 Sol | — | 60.85% | 14.80% | 37.82% |
| 9 | Claude Haiku 5.5 | — | 34.23% | 43.00% | 38.62% |
| 10 | GLM-5.3-Flash | — | 38.35% | 43.90% | 41.12% |
We calculate a blended hallucination rate for LLMs across three benchmarks.
FACTS Grounding judges if models ground their answers in context documents provided for each answer. The grounding score is the share of answers that are accurately grounded. We invert the score to get the share of answers with ungrounded facts, which we interpret as the hallucination rate. The benchmark comprises 1,719 questions split evenly into a public and a private dataset to mitigate the risk of benchmark contamination.
AA Omniscience v1.1 Hallucination Rate measures how often a model answers incorrectly when it should have refused to answer or admitted to not knowing the answer. The benchmark does not provide grounding documents as context and instead measures the parametric knowledge of the models.
HalluHard is a challenging multi-turn hallucination benchmark that covers four domains: legal, research, medical, and coding. It runs models with and without a web search tool. The hallucination rate expresses the fraction of answers where the model fails to provide cited evidence for its claims or the provided evidence does not support the claim.
Our top 10 ranking requires each model to appear on at least two out of the three benchmarks.
Google DeepMind's Gemini 4 Argon has the lowest hallucination rate. It tops the leaderboard at a blended hallucination rate of 15.43 percent. Grok 4.7 sits in second place with a blended hallucination rate of 25.71 percent. GPT 6 Astra is in third place with a blended hallucination rate of 27 percent. It's the only model in the top three that appears on all three benchmarks, so it has the broadest benchmark support.
OpenAI has four models in the top 10 of LLMs with the lowest hallucination rate. Google and Anthropic each have two models on the list.
GLM-5.3 Flash is the best open-weight model when looking at hallucination rates. It hallucinates at a blended rate of 41 percent. This comes with the caveat that open-weight models are rarely benchmarked outside of AA Omniscience.
If only looking at AA Omniscience, there are other open-weight models, such as Meta's Muse Spark or Cohere's Command A+, that have lower hallucination rates than GLM-5.3 Flash.
Does RAG prevent hallucinations?
Retrieval-Augmented Generation (RAG) is a technique that provides the model with retrieved context before it answers a question. It's a common misconception that RAG eliminates hallucinations in LLMs. Benchmarks show that RAG can reduce, but not fully prevent, hallucinations.
On HalluHard, we see a median relative reduction of hallucination rates by 30 percent when comparing results with and without web search enabled. Claude Opus 4.5 gets the biggest uplift from web search grounding: it improves from a 60 percent hallucination rate without search to around 30 percent with web search. It still sits behind GPT 6 Astra, though.
Retrieval architecture also has a major impact on agent hallucination rates. In our research, we show that knowledge graph retrieval outperforms standard vector RAG, and combining the two is even stronger: a hybrid graph-and-vector approach cuts hallucination rates by 22 percentage points compared to embedding retrieval alone.
Do LLM hallucination rates decrease over time?
One might assume that hallucination rates are improving with every model release. Indeed, we see support for intergenerational improvements within model families. On AA Omniscience v1.1, GPT-5.6 Terra has an official hallucination rate of 89.37 percent, improving to 34.50 percent for GPT-6.1 Terra. GPT-6 Sol clocks in at 60.85 percent, while GPT-6.1 Sol scores 51.88 percent. All four results use Max reasoning effort. Claude Opus 5.5 hallucinates at a rate of 50 percent compared to 59 percent for Opus 5, the model's predecessor.
However, some model releases do regress compared to a previous release in the same model family. In the chart below, we can see that Opus 4.6 has a much higher hallucination rate on the FACTS (v2) benchmark than the previously released 4.5 version. Opus 5.5 also regresses compared to Opus 4.8, but it's only a minor difference.
Hallucination rates over time
- GPT Terra · AA v1.1 · Max
- GPT Sol · AA v1.1 · Max
- Claude Opus · FACTS v2
Claude Opus
Swipe or scroll horizontally to see the full chart.
GPT Terra
GPT Sol
View the chart’s source tables
| Model | AA v1.1 official hallucination rate |
|---|---|
| GPT-5.6 Terra | 89.37% |
| GPT-6.1 Terra | 34.50% |
| GPT-6 Sol | 60.85% |
| GPT-6.1 Sol | 51.88% |
| Model | Hallucination rate |
|---|---|
| Claude Opus 4.5 | 37.9% |
| Claude Opus 4.6 | 45.4% |
| Claude Opus 4.7 | 38.4% |
| Claude Opus 4.8 | 30.5% |
| Claude Opus 5 | — |
| Claude Opus 5.5 | 31.5% |
In practical terms, this means that each new model release should be evaluated for the intended use case instead of making the switch blindly.
Benchmarks for LLM hallucinations in specialized domains
PlaceboBench is our specialized benchmark for hallucinations in pharmaceutical RAG. It's a strict benchmark that caters to the high factuality and groundedness requirements for AI applications in life sciences.
GPT 5.4 tops the benchmark at a hallucination rate of 24 percent, closely followed by Gemini 3.1 Flash Lite and Gemini 3 Pro. Claude Opus 4.6 has the highest hallucination rate in the benchmark: 64 percent of its answers contain at least one hallucination.
In a 2025 study, Stanford researchers showed that even purpose-built AI tools hallucinate. The researchers tested legal research tools from Lexis and Thomson Reuters for hallucinations. Both systems had access to data through a retrieval system. Lexis+ AI had access to a wide body of case law and primary sources. Thomson Reuters Ask Practical Law AI had access to articles written by an in-house team of lawyers. The Stanford team reported hallucination rates of 17 percent for Lexis and 19 percent for Thomson Reuters. Both systems have since been superseded by new AI products from the respective companies.
Conclusion
Modern LLMs continue to hallucinate. Given that we attempt to solve increasingly challenging tasks with these models, it is unlikely that hallucinations will disappear completely.
The only way to make a good decision on which LLM to pick for your use case is through your own evals. On top of that, your AI system will likely need additional measures to prevent hallucinations.
Would you like to receive more content like this?
Subscribe to our email list
