AI Hallucination Rates by Model — 10 LLMs Ranked

Every LLM hallucinates. The question isn't whether — it's how often and under what conditions. In this article, we rank 10 leading models by their measured hallucination rates on factual benchmarks, explain the four root causes behind every fabrication, and show which models are safest for accuracy-critical applications like medical summarization, legal research, and financial reporting.

Quick takeaway: Claude Opus 4.8 and GPT-5.5 lead the pack with hallucination rates below 2%. At the other end, smaller open models like Gemma 4 and Gemma 3.5 Flash range from 5–10%. But raw numbers don't tell the whole story — context matters enormously.

Hallucination Rate Rankings — 10 LLMs Compared

Rates measured on the TruthfulQA, HaluEval, and FActScore benchmarks (June 2026). Lower is better. "Risk tier" accounts for both rate and severity of hallucinations.

Rank	Model	Hallucination Rate	Risk Tier	Best For
1	Claude Opus 4.8	≤1.8%	🟢 Very Low	Legal, medical, finance
2	GPT-5.5	≤1.9%	🟢 Very Low	General research, coding
3	DeepSeek V4 Pro	≤2.4%	🟢 Low	Cost-sensitive factual tasks
4	Claude Sonnet 4.6	≤2.5%	🟢 Low	Balanced cost/accuracy
5	GPT-5.4	≤2.8%	🟡 Moderate	Creative writing, brainstorming
6	Gemini 3.1 Pro	≤3.2%	🟡 Moderate	Multimodal analysis
7	Kimi K2.6	≤4.5%	🔴 Elevated	Drafting, non-critical tasks
8	GLM 5.1	≤5.0%	🔴 Elevated	Exploration, prototyping
9	Gemini 3.5 Flash	≤6.0%	🔴 Elevated	High-speed, low-stakes
10	Gemma 4	≤8.0%	🔴 High	Local experimentation only

Sources: TruthfulQA MC2 (June 2026), HaluEval 2.0, FActScore v2. Rates are approximate — individual use cases vary. See full model specs and rankings →

The 4 Root Causes of AI Hallucinations

Understanding why a model hallucinates is more valuable than memorizing a rate. Every hallucination traces back to one (or more) of these four root causes:

1. Training Data Gaps

The model encounters a question at the edge of its knowledge — but instead of refusing, it generates a plausible-sounding extrapolation. This is the single most common cause. Models trained on web-scale data know a staggering amount, but that knowledge has hard boundaries (typically the training cutoff date). When pushed past those boundaries, all models fill the gap with educated guesses.

Example: Ask any LLM for the exact vote count of a bill passed last week, and it will either fabricate a number or wrongly state the bill hasn't passed yet — because the event occurred after its training cutoff.

2. Probabilistic Overconfidence

LLMs predict tokens, not facts. When the probability distribution across candidate tokens is flat (multiple words are equally likely), the model still picks one — and sometimes it picks wrong. This is why lowering the temperature parameter helps: a temperature of 0.0 forces the model to always choose the most probable token, which on factual tasks is more often correct.

3. Prompt Ambiguity

A vague or leading question pushes the model toward fabrication. "Tell me about the 2025 Nobel Prize in Physics and who won it" — if there's ambiguity about which prize, the model might confidently describe a fictional award. The fix is specificity: name the field, year, and detail you want confirmed.

4. Context-Window Drift

In very long conversations (100K+ tokens), models can lose track of facts established earlier in the dialogue. A model might contradict its own earlier statement — not out of malice, but because the earlier context has been effectively "forgotten" as new tokens push it out of the attention window. This is especially dangerous in legal and medical use cases where consistency is critical.

Which Models Are Safest for Accuracy-Critical Work?

If your application involves legal citations, medical summaries, financial figures, or academic references, choose from the top tier:

Use Case	Recommended Model	Why
Legal document review	Claude Opus 4.8	Constitutional AI training penalizes fabrication
Medical literature summarization	GPT-5.5	Best retrieval grounding among frontier models
Financial reporting	Claude Opus 4.8	Lowest hallucination rate overall
Academic research (budget-constrained)	DeepSeek V4 Pro	Competitive accuracy at far lower cost
Code generation	GPT-5.5	Best coding benchmarks + RAG-backed
High-volume, low-stakes summarization	Gemini 3.1 Pro	Good accuracy-to-cost ratio

5 Ways to Reduce Hallucinations in Your Application

1. Use RAG (Retrieval-Augmented Generation): Ground responses in your own documents. This is the single most effective technique — it reduces hallucinations by 40–70% in production systems.
2. Set temperature to 0.0–0.3: For factual tasks, skip creativity. A low temperature forces the model to choose the most probable token, which correlates with correctness on fact-based queries.
3. Add explicit refusal instructions: System prompts like "If you are uncertain about any fact, say 'I'm not sure' rather than guessing" measurably reduce fabrication rates.
4. Deploy factuality guardrails: Use a second pass — either a simpler model or a rule-based system — to verify key claims against a trusted source before showing output to users.
5. Pick the right model for the task: A model with a 1.8% hallucination rate (Claude Opus 4.8) will fabricate roughly 1 in 55 factual claims. A model at 8% (Gemma 4) will fabricate roughly 1 in 12. For high-stakes work, that delta is everything.

Updated June 18, 2026. Hallucination rates based on publicly available benchmark data. Individual model performance varies by domain and prompt quality. Compare all 10 models →

← Back to News