Turn on extended thinking to cut hallucinations 30–60% via in-trace self-correction

Digital Applied·medium signal

2026 hallucination benchmarks find extended/visible reasoning cuts hallucination rates 30–60% across task families, with the mechanism being self-correction during the reasoning trace — the model catches its own confabulations before emitting them. This is distinct from using thinking purely for harder reasoning: the win is factual reliability. Builder move: for fact-sensitive generation (summaries, extraction, claims), enable extended thinking even when the task looks 'easy,' since the trace acts as a built-in fact-check — and reserve MAX_THINKING_TOKENS=0 only for genuinely rote turns.