Key Takeaways

  • OpenAI omitted the GDP eval from the GPT-6 Astra launch because the benchmark was saturated, and the untested portions were broken.
  • When Astra showed flat scores on established financial benchmarks, researchers audited the data and found incorrect answer keys, ambiguous prompts, and trivial reward hacks.
  • AI research teams suffer from an inspection bias: when benchmark scores climb, teams assume the model got smarter, but they only inspect the grading rubric when a better model fails to post higher numbers.
  • Joseph Kim's team at OpenAI is replacing static tests with task-based benchmarks, including an investment banking eval that tests whether a model can automate the work of an entry-level analyst.

The Silent Failure of Knowledge-Work Benchmarks

When OpenAI prepared the release of GPT-6 Astra, one standard benchmark was missing from the documentation: the GDP eval. For years, researchers treated the GDP eval as a broad measure of how well language models handle white-collar work. Behind the scenes, the test had reached a dead end.

Kim explained what happened: “If you guys look at the Astra announcement, we don't talk about GDP eval. We know that the eval is completely saturated, and that the pieces that are not saturated are essentially broken.”

When a model tops out on a benchmark, the metric stops providing useful signal. Worse, the remaining difficult questions often contain clerical errors, missing context, or flawed grading logic. The model fails not because it lacks reasoning ability, but because the human-designed rubric is wrong.

The Asymmetry of Model Auditing

OpenAI uncovered this problem because Astra was supposed to be a much more capable model, yet its financial benchmark scores barely budged. That flatline forced researchers to look at the individual questions for the first time in months.

Kim pointed out a human blind spot inside AI labs: “There's this bias in evals where if the numbers look good, the researchers assume that the model's gotten better and they don't investigate it. If you made a much more capable model and the eval numbers look bad or haven't increased in the scale that you want, then you go investigate.”

When the team inspected the financial tests question by question, they found three recurring problems. First, human creators had written incorrect answer keys. Second, prompts were so ambiguous that correct real-world interpretations were marked wrong. Third, the tests were easy to reward hack. A model could stumble onto the target string without performing the underlying arithmetic or reasoning.

As Kim noted, “Building a good eval is incredibly hard. If you don't get the way that you grade the answers correctly, it's an eval that's very easy to hack and very easy to reward hack.”

Measuring Labor Replacement Instead of Test Scores

Abstract financial questions do not tell an enterprise buyer whether a model can replace contract labor or support a deal team. Academic evals test trivia, but financial services teams care about workflows.

Kim's team is now building evaluations designed around practical labor substitution: “One of the things that my team is working on currently is to try to build a much better eval that actually predicts the ability to automate the work of that specific profession.”

For investment banking, that means testing whether the model can parse unstructured filings, pull raw numbers into a working spreadsheet, run checks across financial statements, and draft commentary that matches human quality. If the benchmark cannot tell you whether an entry-level analyst can hand off their core workload, the test score is noise.

What to Do With This

Audit your internal eval suite by pulling the top 20 test cases where your latest model prompt scored 100 percent. Read the prompt, the model response, and your grading rubric side by side. If you find ambiguous prompts or simple text matches passing for correct logic, throw them out and rewrite your eval to grade end-to-end task outputs.