Lots of good stuff in this thread, catching up on the comments after Google IO craziness.
A few reactions, comments, and clarifications:
- I hear that you all want raw thoughts, the value is clear, there are use cases that require them, and seems reasonable to want them in the API as well
- Why be excited for summaries? The raw thoughts have been disabled in the API for a while now, so summaries was a step closer to allowing devs to programmatically at scale see what the model is doing in reasoning steps. I am still excited about this because there are certainly use cases that will benefit from it (but as folks have pointed out, clearly not all use cases)
- There were a few comments on Gemini 2.5 Pro 05 06 being a regression for certain tasks, we have seen lots of reports of these and are working on some improvements!
- There were some comments on model quality regressions with thinking summaries. To be clear, at the API level, for the last few months, we have not had raw thoughts in the API. The summary is purely cosmetic in that it’s returned after the fact and does not impact model performance in any way, the model still reasons with full thoughts for all queries
- On the Gemini app side, I do personally think this is a real feature improvement. Gemini app is being built as a universal assistant. The % of people who will or do read thought thoughts in the Gemini app is very small, the better product experience on the margin is not showing them (and instead showing summaries, especially as models do more and more in a single pass)
- On the AI Studio side, I hear you all on wanting to have raw thoughts back, it is something we can explore. The dev use case makes perfect sense here and is the reason we left the raw thoughts in AI Studio to begin with when we removed it from the API
- In the short term, we will keep working on ways to make summaries better and see how to expose more
- In the long term, as models do more in the reasoning steps (tool use and otherwise), I can easily imagine that raw thoughts becomes a critical requirement of all AI systems given the increasingly complexity and need for observability + tracing
- For folks using thoughts for function calling / tool calling, we just added Thought signatures which lets you pass an encrypted thought representation in multi-turn conversations that require tool calls: Gemini thinking | Gemini API | Google AI for Developers
- Thought summaries in the API are experimental, specifically so we can get feedback and influence the roadmap, so thank you to those who shared cases where they don’t work well for specific use cases!
I hope this helps!