OpenAI claims ChatGPT’s new default model hallucinates way less

The headline reduction figure is self-reported against OpenAI's internal SimpleQA-style benchmarks, and the announcement does not specify which task distribution, domain, or prompt format was tested. That omission matters enormously: hallucination rates vary by an order of magnitude depending on whether you're testing factual recall, multi-hop reasoning, or numerical claims.

This announcement lands in the same week as the ARC Prize Foundation analysis (covered May 2nd) showing GPT-5.5 still fails on three systematic reasoning error patterns despite scale. Those two data points sit in direct tension: OpenAI is claiming factual reliability gains while independent researchers are documenting persistent structural failure modes in the same model family. Also relevant is the goblin-training incident from May 1st, which demonstrated that reward signal misconfiguration can produce widespread behavioral artifacts that evade internal testing. That precedent gives legitimate grounds to ask whether the evaluation suite used to measure this hallucination reduction is itself well-specified.

Watch whether GPQA Diamond or BioASQ third-party leaderboard scores for GPT-5.5 Instant, posted by independent evaluators within the next 60 days, show comparable factual accuracy gains. If they don't, the internal benchmark is likely measuring a narrow distribution that doesn't generalize.