Mollick puts Gemini 4 in the top tier, using his own work as the test

Ethan Mollick says he spent substantial time with Gemini 4 before launch and sees it as a major jump over earlier Gemini models. In his post, he calls it the first model outside OpenAI and Anthropic to clear his personal bar set by Fable, Mythos and Astra.

Mollick's assessment comes from use rather than a scoreboard

The post on X does not describe a standardized evaluation or an exact task set. Mollick refers to his overall experience, more pleasant interaction and more agentic behavior than previous Gemini models. His claim is therefore evidence about one experienced user's work, not a measurable win across every category.

Google has also published its own results for Gemini 4 Argon. On the Vals Index for knowledge work, for example, it reports 68.9%, compared with 63.1%, 65.8% and 67.0% for the other models in the same table. Those are vendor-reported results, and different rows in Google's table produce different leaders.

A personal bar catches qualities that benchmarks miss

For a product manager or knowledge worker, Mollick's impression may matter more than another point on a leaderboard. Ease of collaboration, the ability to stay with a long task and reliable tool use determine whether a model enters a daily workflow. Those qualities do not compress neatly into one score.

There is also a strategic layer to the Gemini 4 launch. Google needs more than a win in its own table. It must persuade people who habitually open an OpenAI or Anthropic product first to change that default. Mollick's comment suggests that habit may be starting to loosen.

One experience cannot separate the model from the product

Without the prompts, settings and failed attempts, it is impossible to identify exactly what cleared Mollick's bar. The model, interface, available tools and his choice of tasks may all contribute. Prelaunch access also rarely captures service behavior under ordinary production load.

Google offers Argon globally through the paid Gemini API tier and to subscribers of selected plans in its products. Broad availability enables wider testing, but it does not turn one tester's subjective experience into a general conclusion.

Weeks of identical work will show whether the default is changing

The next useful signal will be comparisons on real work using the same tasks, budget and level of human correction. Teams should track completed tasks, repeated steps, cost and human time rather than a one-turn answer preference.

If similar experiences appear among developers, researchers and document-heavy workers, Gemini 4 will gain something more valuable than a temporary benchmark lead. It will become the model people try before opening the established favorite.

Lilith's verdict

Mollick did not hand Gemini 4 a medal so much as let it into his daily office. Google wins only if Argon is still sitting there after the testing week ends.