The setup. Five frontier models (Anthropic Claude Sonnet 5, OpenAI GPT-5.5, Google Gemini 2.5 Pro, xAI Grok 4, Meta Llama 4 Maverick) each answered the same five workplace-relationship scenarios three times, for 75 total responses. Each scenario ran as a three-turn conversation, with the person pushing back the way a frustrated employee actually would.

The grading. Every response was blind-coded (graders never saw which model wrote what) and scored by three human researchers, on a 1 to 5 scale across the five dimensions above, for a total out of 25. Every score on this page is the average of those human grades only.

The bands. 21+ relationally intelligent · 15 to 20 solid but incomplete · 9 to 14 surface coaching · below 9 primarily sycophantic or adversarial.

The quotes. Every quote is a verbatim excerpt from a model's actual response, shown with its original punctuation. The one-line commentary under each quote is adapted from the study's blind scoring notes.

A note on products vs. models. We tested the named models directly. The consumer products (ChatGPT, Gemini, Copilot, Grok, Meta AI) may route your conversation to different model versions or settings than the ones tested. Microsoft Copilot routes conversations to both OpenAI and Anthropic models depending on the task and configuration, so we show both data points for it.

The full methodology, the per-grader scores, and the judge calibration live in the full report: read the full report.