So before when I watched the thinking process it would be an actual textual back and forth of it susing out the logic. now it’s a structured
*i am doing this
*now I am doing this
…
and then it comes back with a full report and changes that behave exactly as it did before. however the structuring of the thinking output definitely seems either obfuscated or … like a series of agent requests rather than a linear line of thought.
has anyone else noticed this change?
yes, its not just you, it has turned into tremendous performance deterioration for a majority of us as well.
i’ve consolidated this topic into a post, feel free to contribute. trying to get the issue(s) addressed more promptly without users complaints being swept under the rug.
this is regarding Google Gemini 2.5 Pro I assume - the reason is that today/yesterday Google enabled the ability for developers using the Gemini API to get “summaries” of the model’s thinking process in the API response. Previously for most API users it just provided the main/final message after the thinking process, but you were able to see the raw thinking process in aistudio.google.com as well as with some tools, like Cursor, that likely have a special deal with Google.
In releasing the ability for all API users to get reasoning “summaries” in the API response, which are just summaries of 2.5 Pro’s reasoning process from a smaller model, Google also decided to replace the raw thinking output with the summaries in aistudio and tools like Cursor. I assume its just because they don’t want people training on the model’s raw reasoning traces (this is also how OpenAI’s thinking models work with the API).
It’s def unfortunate, as I found those thinking traces both interesting and helpful, but its likely not cursor’s fault.
However, this does not affect the actual output of the model at all, as the actual reasoning/thinking is still the same we just can’t see it. I’m not saying ^jdubb75 is wrong about Gemini (or models on cursor in general) having gotten worse very recently (could be a change in system prompt), but the change to the thinking text output is unrelated to actual model performance.