It is clearly much better, the thinking is sharp and clear, the concepts are much more refined and specific. The rationale is much stronger.
Thank you
19 Likes
āModified by moderatorā Google removed 0325 pro when the new 05 model came out, secretly hiding the 0325 and directing to 05 unannounced.
But when the light weight (cheaper) flash came out , they kept the old flash. If you donāt have the capacity to host, at least have the gut to admit it like open ai.
Donāt the google poeple who are reading these threads feel embarrassed about your company?
12 Likes
exactly, they let the old flash stay but not the old pro. no reason whatsoever. it makes no sense
9 Likes
It makes good sense for hyping up the stock price for the general public with 0325, then secretly hiding it because the ācost-benefitā for Google commercial interest, leaving developers suffering.
Open source is the future
10 Likes
Personally I agree 03-25 follows instructions precisely and the rationale is awesome. So dissapointed itās discontinued how they can bring it back. 
5 Likes
Same this just organizes stuff but the older version has more freedom in its thought process
5 Likes
Hey Guys,
Could you please provide more details on the specific task where the 2.5 Pro (03-25) model outperformed the current 2.5 Pro model? While we have no plans to reintroduce the previous version, understanding these performance differences is important for identifying areas to improve in future iterations.
Thank you!
If I can chime in, a big benefit of the old version was the direct and full CoT (Chain of Thought) and the ability to steer the models CoT using set frameworks/prompts that would lead to longer CoT (time spent āthinkingā), that adheres to a specific format/structure when reasoning/planning. This would lead to very nuanced, detailed, and long outputs. Currently, the CoT is an approximation of each āstep/section.ā
This was marked as closed, but this dev question was left unanswered, and I wanted to at least provide this specific feedback as it applied to my use cases and what made that model feel more robust and powerful, providing longer outputs that were more nuanced and well-reasoned. This was specifically useful in creative writing tasks when a specific framework is in place to follow.
I bring this up because both Sonnet 3.7 and this version of 2.5 pro (as well as the initial 2.5 flash version still available) allowed more direct control of their CoT process, allowing CoT-specific system prompts to exist and function. These same prompts donāt produce the same intended CoT or specific outputs, usually leading to the steps intended to happen during the models CoT either spilling into the final output or only being present in the final output. Sonnet 4.0 and the latest versions of 2.5 stable (both pro/flash) where CoT is summarized have all regressed in this way or in terms of the usersā ability to steer the modelsā thinking and outputs.
I can just say that the response from 03-25 is very sharp, it can pinpoint to a very specific point in concept with high precision.
I think you might compare internally, the response of 03-25 and newer models, on the conceptual refinement, clarity, specificity. It can discern logic and concepts with high resolution.
2 Likes
Thank you - your feedback is invaluable as we work to continuously improve the Gemini API experience.
Maybe you can ask 03-25 and latest model same questions, and feel it yourself, on the aspects of conceptual clarity, resolution, precision, nuanceā¦sorry for not being specifc enough about it
1 Like