Gemini 3 Pro Preview: Not Quite Baked

It was an open secret that Gemini 3 Pro was going to drop this week. OpenAI released GPT 5.1 as barely a fresh coat of paint over GPT 5 to get out ahead of the news, and xAI launched Grok 4.1 without API support.

Then Google themselves released a Gemini 3 Pro preview that is, in fact, not fully baked. 

First, the good news. Even with its issues, Gemini 3 Pro is close to the top of the power ranking, just barely behind GPT 5.1. But fixing those issues could put it convincingly ahead, and also reduce the cost difference that makes it a tough choice for now.

We observed two kinds of problems while testing Gemini 3 Pro Preview. The first is that it can get stuck in a repetitive loop in a way that feels more like a tiny 3B model than a flagship entry from the big three labs; this makes it both slower and more expensive to run. We observed this several times, e.g.

    *   I'm done.

    *   I'll submit.

    *   I'll submit.

    *   I'm done.

    *   I'll submit.

    *   I'll submit.

    *   I'm done.

    *   I'll submit.

    *   I'll submit.

    *   I'm done.

    *   I'll submit.

    *   I'll submit.

    *   I'm done.

It continued like this for hundreds of lines.

The other problem we saw is that about 10% of the time, the server will dump reasoning tokens into the text stream instead of separating them out, resulting in no reasoning tokens at all and a confused text stream. (This is especially a problem when you're trying to use structured output.)

We look forward to retesting Gemini 3 Pro when these issues are resolved. In the meantime, you can check out the rest of the November Power Ranking here.