> So, so weird that they still don't want you to see their models' reasoning process

It's not weird at all. R1-distills have shown that you can get pretty close to the real thing with post-training on enough completions. I believe gemini has also stopped showing the thinking steps (apparently the GLM series of open access models were heavily trained on gemini data).

ToS violations can't be enforced in any effective way, and certainly not cross-borders. Their only way to maintain whatever moat thinking models give them is to simply not show the thinking parts.

Google actually switched to an OpenAI system for 2.5 Pro's Chain-of-Thought yesterday on the Gemini app and AI Studio ("I did this; I did that. etc"). Apparently it still shows via API, but no clear how long. Also, in my experience, if you select the "Canvas" output, you still get the old style CoT.

And yes, the above is true even if you are ULTRA.

You can still view your old thinking traces from prior turns and conversations.

My heart just broke to hear this. Although I honestly don't read the thinking output very often. But I had been cheekily copy-n-pasting the info for my own records.

I agree, but there's always Deepseek. They're publishing and open-sourcing more than anyone these days.

Not only that. I have an agent product and I’m currently blocked from using their reasoning models on Azure for having asked for a chain of thought, which apparently is against the ToS.

The customer service itself was surreal enough that it was easier just to migrate to Anthropic

Is this true? I can click open o3’s dialogue and see a running monologue. I guess it might be a summary of the actual reasoning though.

Correct, you are not seeing the reasoning chains.

I may be giving Gemini too much credit, actually - seems like its "reasoning" may be a summary as well.

They changed it yesterday or so: it used to show the actual reasoning, now it no longer does. And the reasoning was quite useful to see if it was going down the wrong track, the summary is much less so.

That's disappointing. I was getting a lot of utility from reading through the thoughts returned by Gemini when I used it in Cursor - occasionally even learning something new from its stream of "consciousness". Obfuscating the information because it can be used to train competitors seems misguided, if understandable.

Agreed. Right now deepseek’s R1 has uncensored stream of consciousness in open weights. I think it’s interesting that teams feel the streams should be proprietary. They must be doing something a little different than R1, or it wouldn’t be worth the extra engineering work.

It is a summary