Ethan Mollick says Gemini 4 passes his top-tier bar but lags in multi-agent work

Ethan Mollick says Gemini 4 is the first model from a company other than OpenAI or Anthropic to pass his top-tier bar, in an X post on October 13, 2026.

He says it is a big jump over earlier Gemini models, matches rivals on creative writing, and lags on complex multi-agent work. He also says it tends to be overly certain and a little lazy. These are his observations from one test, a city-builder prompt run on all models with Gemini 4 in the Antigravity harness. His post includes no comparative scores or other test data.

In a follow-up post the same day, Mollick says using Gemini 4 makes him think many other models distill Claude or GPT or use a similar training pipeline. He says Gemini 4 feels quite different from other models.