I have a weakly held conviction (because it is based on my personal qualitative opinion) that Google aggressively and quietly quantizes (or reduces compute/thinking on) their models a little while after release.
Gemini 2.5 Pro 3-25 benchmark was by far my favorite model this year, and I noticed an extreme drop off of quality responses around the beginning of May when they pointed that benchmark to a newer version (I didn't even know they did this until I started searching for why the model degraded so much).
I noticed a similar effect with Gemini 3.0: it felt fantastic over the first couple weeks of use, and now the responses I get from it are noticeably more mediocre.
I'm under the impression all of the flagship AI shops do these kinds of quiet changes after a release to save on costs (Anthropic seems like the most honest player in my experience), and Google does it more aggressively than either OpenAI or Anthropic.
Gemini 2.5 Pro 3-25 benchmark was by far my favorite model this year, and I noticed an extreme drop off of quality responses around the beginning of May when they pointed that benchmark to a newer version (I didn't even know they did this until I started searching for why the model degraded so much).
I noticed a similar effect with Gemini 3.0: it felt fantastic over the first couple weeks of use, and now the responses I get from it are noticeably more mediocre.
I'm under the impression all of the flagship AI shops do these kinds of quiet changes after a release to save on costs (Anthropic seems like the most honest player in my experience), and Google does it more aggressively than either OpenAI or Anthropic.