BALTIMORE — The second time around, the AI didn’t get smarter. It just got more expensive to run. That is the blunt finding from a University of Maryland Francis King Carey School of Law study released this week.
Researchers gave OpenAI’s newest model, GPT-5.5, the same law school final exams their predecessor took a year earlier. The older model, OpenAI’s o3 reasoning model, had scored anywhere from an A+ to a B, graded on the same curve as human students.
The new model, despite being billed as an upgrade and given extra computing power, did not beat that score. In some areas, it may have done worse. “This apparent plateau, despite a better model and using extra-high inference-time compute, may reflect trends observed in other legal benchmarks for AI and makes theoretical sense,” the authors wrote.
That sentence, buried in the study’s conclusions, is the real headline. Last year’s results had turned heads.
An AI scoring an A+ on a law school exam felt like a threshold moment, a signal that the bar exam itself might be next. The new findings suggest something else: a ceiling, at least for now. The researchers are calling it a plateau.
Not a collapse, not a failure, but a flat line. The technology is still learning the law, they say, but it may be hitting a wall. What changed between the two tests?
The model changed. The compute power changed.
The exams themselves did not. The same team ran the same experiment with the same grading curve. The only variable that mattered was the AI’s own performance, and it did not improve.
For anyone watching the steady drumbeat of AI announcements — each new model promising to be faster, smarter, more capable — this is a sobering result. The straight line of improvement many assumed would continue has, at least in this narrow test, bent flat.
The study matters beyond law school. The researchers are not just testing whether AI can pass exams. They are testing whether it can reason about legal problems the way a human lawyer does.
A plateau on a test suggests a plateau in reasoning, at least for now. What this means for patients, for families, for anyone whose legal fate might one day be touched by AI: the technology is still learning the law, and it may be hitting a wall. That is not a prediction of failure.
It is a description of the present. The researchers plan to keep running the experiment in future semesters, tracking whether the AI’s performance evolves or stalls.
The TaxProf Blog covered the findings, and the discussion among legal educators has been active. For now, the message is clear. The AI did not get worse.
It just did not get better. And in a field where each new release is supposed to be a leap, standing still is a kind of news.
{“@context”: “https://schema.org”, “@type”: “NewsArticle”, “headline”: “Study: Newer AI Model Costs More but Fails to Improve Law Exam Scores”, “description”: “BALTIMORE — The second time around, the AI didn’t get smarter. It just got more expensive to run. That is the blunt finding from a…”, “datePublished”: “2026-06-29T13:16:00+00:00T09:30:00+00:00”, “author”: {“@type”: “Person”, “name”: “Merry Gel Sigui”}, “publisher”: {“@type”: “Organization”, “name”: “InfoPulse Today”, “url”: “https://newsfeed.infopulsetoday.com/”}}
