An OpenAI/deepsense.ai plugin study reports an 88% answer pass rate for GPT-6 Luna on 108 BLS-data questions under Codex. This tests a particular model, plugin and harness together, rather than general model accuracy.
Tasks: Held-out production test split: 108 single-turn questions across 56 BLS surveys.
Limits: The evaluation was reported by collaborators who built or studied the plugin, including OpenAI. It is not an independent external validation of OpenAI models. The article names GPT-6 Luna without identifying the evaluated API snapshot, weights, reasoning setting or run date. Association with the named catalog model does not establish the behavior of a currently served endpoint. The dataset is a private held-out production split. Its code and data are not supplied; it is separate from the runnable five-question synthetic example. The reported pass rate uses an LLM judge and is rounded. No confidence interval, per-question rubric, numeric tool/token budget or full production prompts are available. Three repeated Codex runs report 88%, 89% and 86%; 23 of the 108 questions change pass/fail outcome across those runs. That spread is not a confidence interval or 324 independent test questions. This measures the combined model, plugin and harness on BLS-data questions. It is not comparable to MMLU, HLE, coding benchmarks or a general accuracy score; the article also finds harness-dependent results.
OpenAI Developer Docs / OpenAI and deepsense.ai contributors
