The intuitive assumption is that a model which reasons through a problem step by step should be more reliable, not less -- more deliberation should mean fewer mistakes. This year's benchmark evidence says the opposite on plain factual accuracy: reasoning models exhibit higher hallucination rates than non-reasoning models on the same class of question, a pattern that's shown up consistently enough across independent studies this year to stop reading as noise.
The magnitude is not subtle. One widely-cited 2026 study found accuracy dropping from 98.2% to 64.4% for GPT-4o and from over 90% to 14.4% for DeepSeek R1 under conditions designed to stress-test factual reliability rather than reasoning capability specifically. Separately, a 37-model benchmark put hallucination rates in the 15-52% range depending on task, and domain-specific testing found hallucination rates of 64.1% in medical case summaries and 58-88% in legal research queries -- specifically in citation generation, which is exactly the sub-task where a wrong answer is hardest to catch by eyeballing the output.
The one mitigation that shows up consistently across studies isn't a different model family, it's a different inference setting: extended thinking roughly halves hallucination rates across the models tested -- GPT-5.5 Pro from 8.3% to 4.2%, Claude Opus 4.7 from 9.4% to 5.1%, DeepSeek V4 from 12.7% to 10.4%. That's a real, replicated effect, not a marketing claim, and it means the actual lever available today is a runtime setting (how much thinking budget a request gets), not a model-selection decision alone.
The operational implication is uncomfortable for anything optimizing purely for "smartest model available": a model chosen for its reasoning benchmark score is not automatically the model you want generating a citation, a medical summary, or any output where a fluent, confident, wrong answer is the actual failure mode to defend against. That's a case for output verification and citation-checking as a standing step in a pipeline, not an occasional audit -- and it's a reason to default extended thinking on for exactly the output types this benchmark work flags as highest-risk, rather than treating it as a cost to economize away.
Choosing a model by reasoning-benchmark score alone can make factual reliability worse, not better -- extended thinking is the one mitigation with consistent, replicated effect (roughly halving hallucination rates across model families), so treat it as the default for citation-, medical-, or legal-adjacent output rather than a cost to trim, and keep a real verification step in the pipeline regardless of which model is generating.