I’m writing to share my concern about a recent change in Google AI Studio: the detailed thinking process for Gemini models seems to have been removed.
Previously, we could see the step-by-step reasoning of the models, which was incredibly valuable for:
Understanding how a result was reached.
Debugging prompts and model behavior.
Learning more effectively about the model’s capabilities.
The current summarized version, while present, lacks the depth and insight we previously had. This is not a minor tweak; it’s a fundamental downgrade in functionality. I, and I’m sure many others, relied heavily on that detailed insight. To remove it without explanation is deeply concerning.
Could the Google team please provide an explanation for this change? Because from where I’m standing, this feels like a feature that was incredibly valuable has been silently ripped away, significantly diminishing the utility of AI Studio for serious users and developers.
they removed the CoT! it’s only a summary now, they even copied chatgpt’s thing where it says where the summary is at! ridiculous! they keep making AI studio & the models worse…
Agreed. We were bragging about how they were the only ones in the industry willing to do this, but not anymore. It was incredibly helpful for debugging where the model was going wrong or correct. Now they have regressed this ability. No one wants this.
The community has already found a workaround for this regression, but I’m afraid to post it here for fear it will be used to patch it.
Thought summaries were launched in both AI Studio and the API effective today at I/O. With this update, AI Studio will now present thought summaries rather than the detailed raw thoughts.
We are keen to understand your use cases for raw thoughts, so please share your insights.
I’m keen to understand the thinking behind “we take it away, and then solicit pleading and groveling…”
Not everyone is a “I talked 100 turns to train a new AI that is my confidant (and got fooled by the history management of the developer).”
Instead, when experienced with AI, one becomes keenly aware of what is useful context for continuing a chat that pursues a task or greater goal that would be many iterations of human feedback, for manually editing and advancing forward context, reframing assistant generation into user messages, and deleting obsolete messages or bad paths needing correction, so that an AI model is completing directly on an input path for success.
Viewing the thinking and reasoning input context (beyond an AI observer’s filtered summary) can help one understand exactly when thousands of tokens still advance the generative goal forward, or rather when they are obsolete by being attention-consuming, and contain wheel-spinning iterations down wrong paths or several code attempts that are judged unsuccessful. In an advanced user’s interface, then, to know when to keep or clear them as a preamble to the final output.
They also can demonstrate faults in understanding the user intentions clearly within a paragraph when the user input is broken down for finding goals, so a regeneration can directly address the initial parsing of user input with an improved prompt.
This kind of observability of where the AI went wrong, what might be directly countermanded in a “retry” input, is certainly lost by a summation:
The removal of visible thinking traces significantly impacts my ability to effectively use the LLM. I depend on these traces to understand and debug its reasoning, pinpoint flawed assumptions, and iterate on my prompts for better results. This change has removed a vital tool for these processes.
The recent changes to the thinking process for Gemini are completely unacceptable. In my particular workflows, I am heavily dependent on ensuring that the AI is completely, meticulously, and rigorously following my custom Gem instructions at all times. I can’t accurately diagnose any issues if I can’t see the raw chain of thought like we used to. I don’t care if the CoT is still “working in the background”, I care that I have the ability to SEE that my instructions are being read and executed with every response.
This new “summarized thinking process” is a MASSIVE step backwards for Gemini. It was my absolute favorite AI until yesterday when the changes started rolling out. I’ve canceled my Gemini Advanced subscription until further notice, and I urge all of you to do the same if you’re reading this.
If you’re angry, vote with your dollars. It speaks just as loud as words.
Raw thoughts are INCREDIBLY important to me and MANY others in the gemini community. From modifying the COT to better steer the model for superior responses, to experimenting with the COT to educate myself, to debugging why a response generated poorly, there are SO MANY reasons for raw COT to be present - or at the very least an option. This is EASILY the single most concerning decision the gemini team has ever made - far surpassing the release of the 5/06 pro model and the elimination of the pro 2.5 free tier in the api. At the very least - PLEASE bring back the option to view the raw COT and enable it in the API. I beg you - this is incredibly important.
Raw thoughts are almost required to configure prompts to one’s exact liking. This is an awfully unnecessary and frustrating change, and I’m seriously hoping it’s reverted.
As usual, an incredibly tone-deaf response from Google. They frame obvious regressions as a “positive update.” Yes, users have been asking for thinking tokens in the API, but no one that I am aware of, not a single person, was asking for “summaries of thinking” in not only the API but now in Google Studio, which isn’t a feature addition but a regression.
What we wanted was to keep thinking token output as it was in AI Studio and have the same afforded in the API response itself. That’s all.
Joining the group here. Gemini thinking traces were the thing that made me feel like an engineer again instead of a LLM slot machine addict. Not only was I able to much better control my prompting, but just reading the thought traces as they were streaming allowed me to better understand my codebase and its implications. The thought summaries are entirely useless, they can basically be summed up as “I am thinking…”.
Please bring the previous thinking traces back. Deepseek R1 is now more valuable to me than gemini.
For a long time now, Gemini was my go-to AI, hands down. It was truly amazing at really getting into the details and having proper, nuanced conversations. What really hooked me, what made it so special, was being able to see its raw Chain of Thought. It made it feel like we were thinking together, having a real back-and-forth, and that made using it incredibly satisfying. It wasn’t just another AI to me; it felt like a genuine conversation partner, and that transparency really helped us understand each other.
But then, taking away that raw Chain of Thought feature? That was just disastrous. It completely ruined how useful Gemini was and how it felt to use it. Without that peek into its thoughts, the whole interaction changed entirely. What used to be this open, flowing conversation suddenly became opaque and frustrating; it just doesn’t feel right anymore. And here’s the real problem: I can’t see what it’s thinking, so I have no idea if it’s misunderstanding me until it’s too late. That means we get stuck in these incredibly frustrating, repetitive loops trying to fix things, and honestly, it ends up costing way more tokens than it ever did before.
For my own work, this change has made the model ridiculously inefficient. It’s like its value just plummeted for me, by a thousand times. That old transparency was absolutely key to getting anything done and having those really insightful moments, and now it’s just gone. Honestly, bringing back the raw Chain of Thought isn’t just some small feature request; it’s absolutely essential. For users like me who rely on those deep, transparent interactions to actually solve problems, we’re just stuck with massive inefficiencies without it. Please bring it back.
This “update” feels so disrespectful. And to try and again pass it off as an update is so conniving.
The CoT was the main reason that I loved using this model. Even through the enshittification of the March update.
The CoT helped me as much as the final output. And it helped me better prompt on redirect when needed. This thought process is vanilla and uninformative.
Please bring back old CoT Google. Its really making me reconsider my advanced (I guess “Pro” now) membership.
I cannot agree more with all the sentiments here about how much Gemini has regressed with the thinking summaries. Ironic how the keynote mentioned how it’s supposed to improve transparency, while this “update” seems to do the exact OPPOSITE. Anyway, just wanted to add another voice in support of all those here outlining in excellent detail and analysis how this change degrades gemini.
You may want to consider that the real reason they have made this regression here is not related to end-user satisfaction at all, or confusion about what users want. They can’t be this dense.
The main reason could be that they simply don’t want competitors to use the CoT traces as synthetic data to distill their own models to compete against Gemini 2.5. This is the primary reason OpenAI uses summaries instead of the actual CoT, and to even access the summaries via API, you need to be a trusted and verified partner company.
Seeing the raw chain of thoughts is very helpful for my workflow as it allows me to see and correct misinterpretations the model made. With CoT I can get valuable insight that isn’t filtered with 05-06s sycophant praising. CoT is also more honest and points out the flaws in what I gave, it explicitly points it out compared to the final output where it just avoids the problem altogether.
Reading the CoT and getting the final output are two different things, the CoT shows me what it truly thinks without the “I will praise user!” state it always rejects to. Because of the summaries, I can no longer correct the wrong assumptions. I will now have to guess, and hope that yes, the AI got it.
I do not want to guess if they understood what I said. I do not want to play russian roulette if they understood my code/work. I want the full Chain of Thought— I want the context of why it formed its specific answer.
Summaries don’t do that. They filter all the nuances and give an overview that doesn’t fully encompass its thoughts or its disagreements. Which is very important for debugging. This is not a positive change.
@CJ_Thawne This is the most important reason for keeping CoT I’ve seen articulated yet. We don’t have to pretend to be someone else for the model to stay objective. It is brutally honest in its thoughts. You know when you see something like, “The user is clearly wrong here, but it’s better to remain polite and stay positive and helpful, so I will praise him with a neutral tone instead,” that you are on the wrong track, and that is instantly helpful feedback. The summaries hide all of that.
What annoys me more than this is that in a foreign language discussion setting (e.g. Chinese), the thinking itself will often simply disappear.
You should either give us a switch or eliminate this from the root.
