GPT-5.5 flags FrontierMath benchmark errors

The irony is that benchmark QA is now reaching the same capability tier as the benchmark it is QAing. This is primarily a benchmark integrity story, not a new model release or product launch. If the reported error rate holds, corrected FrontierMath scores could shift how we read recent math-capability claims. GPT-5.5 being useful as a sanity-check tool suggests model-assisted eval review is becoming a practical workflow, not just a research curiosity. The broader signal is that hard benchmarks now need stronger review pipelines because models are getting better at spotting flaws in the benchmarks themselves.