Anthropic's October 16 research describes a low-cost way to monitor model misuse: after a response is generated, the model assesses its own output while reusing its KV cache. The company reports about ten additional generated tokens and less than 2% incremental inference compute. On its Claude Fable 5 evaluation, it found 78% of the cases identified by a prompted classifier and 94% of that classifier's high-priority cases at a 1% false-positive rate. The denominator is another classifier's labels, not all true misuse. Anthropic says this is an offline monitoring technique deployed since June across a large share of traffic; it is not a claim that dangerous actions are stopped in real time. The CTO lesson is to measure supervision cost and detection against an explicit reference, then test it on one's own failure modes.
OpenAI's October 19 work on coding in low-resource programming languages addresses another evaluation trap. Its internal benchmark asks models to build a SQLite-compatible engine across 20 languages; the company reports average pass rates of 87.4% for Astra and 26.8% for GPT-5.6 Sol. On a separate COBOL migration test, it reports 86.6% observable parity for GPT-6.1 Sol at low reasoning effort versus 5.4% for the earlier model, with 97.9% at max effort. These are vendor-run tasks with their own test design, not a forecast for a bank's COBOL estate. A compile-clean migration can still change rounding or truncation and therefore change cash flows. The useful next step in a financial firm is hidden, domain-specific parity testing of outputs and edge cases, with cost measured alongside pass rate.
Sources: Anthropic, “Scaling Misuse Detection with Self-Assessment,” October 16; OpenAI, “Advancing coding in low-resource programming languages,” October 19.