Stanford study tracks ChatGPT drift: GPT-4 math and coding shift over months

For months, developers have complained that GPT-4 feels noticeably dumber, while OpenAI insisted nothing was wrong. Now, researchers have brought receipts.

A team from Stanford and UC Berkeley evaluated the March and June 2023 API versions of GPT-3.5 and GPT-4. Across tasks including math, code generation, sensitive questions, and medical tests, they found that the exact same model endpoints shifted dramatically in just three months.

Math tanked and code formatting broke

The most dramatic swing showed up in basic math. In identifying prime numbers, GPT-4 dropped from 84% accuracy in March down to 51% in June. Ars Technica noted that on certain specific subsets, accuracy cratered from 97.6% to a measly 2.4%.

Researchers traced part of this plunge to a sudden drop in GPT-4's willingness to follow chain-of-thought prompting. Its ability to follow user instructions degraded, which became a common thread across multiple regressions.

Code generation took a hit as well, with both GPT-4 and GPT-3.5 introducing more formatting errors in June than March. Meanwhile, GPT-4 became much more guarded, refusing to answer sensitive questions and survey prompts.

Curiously, the drift did not move in one direction. While GPT-4 improved on multi-hop questions, GPT-3.5 got worse. And on prime numbers, GPT-3.5 actually performed significantly better in June than it did in March.

OpenAI pushes back, researchers debate

The findings gave fresh ammo to internet theories, from model distillation to save GPU compute to safety fine-tuning breaking other reasoning tracks.

OpenAI VP of Product Peter Welinder pushed back hard, tweeting that the company has not made GPT-4 dumber and that new versions are smarter. His counter-hypothesis was that as people use the tool more heavily, they simply notice flaws they previously missed.

Princeton computer science professor Arvind Narayanan also urged caution regarding the study's conclusions. He pointed out that the coding benchmark evaluated whether code was immediately executable rather than conceptually correct, arguing the shifts resemble side effects of fine-tuning rather than an objective decline in intelligence.

Why it matters

If you build software on top of commercial LLM APIs, this study is a flashing red light. A production pipeline that worked reliably last week can silently break overnight because the vendor tweaked a system prompt or post-training weights without warning.

For the AI industry, it exposes the biggest flaw in the current black-box API business model. Vendors want enterprises to treat LLMs as mission-critical infrastructure, yet they manage them like consumer web apps with silent A/B testing and vague release notes.

We think this study's real value is not whether GPT-4 got objectively dumber, but how thoroughly it exposed vendor opacity. Model drift is real, and until providers commit to rigorous changelogs and frozen versioning, developers will have to spend significant resources continuously monitoring the models they pay for.