Sounds like some is behind.

Yeah that is why I said "apparently".

Though I will say I have given it a try now and it is damn good. And 2.5 has been so great that I'm willing to give them the benefit of the doubt. That is not something I would have given Google just a few short months ago. The wind has very much shifted.

Are there benchmarks for deep research?

Yeah the UI is a little messy. Looks like they initially thought to shove everything in the model selector. Now they are changing that so the operation is selected separately to the model.

You select 2.5 then press deep research.

😆 obligatory xkcd about standards.

Given how Excellent 2.5 has been I believe it. My advanced subscription I got with my phone is finally starting to pay off.

I still use google a lot. It is a totally different purpose to language models.

Slides and docs doesn't use Gemini 2.5. in fact I think it used 1.5

Yeah stdio, especially for local use, makes sense. 

But I'd like to use stdio in a request/response pattern so that it extends naturally to the REST interface. It feels a bit like they did the reverse.

I just saw another thread complaining about how expensive it is to run an MCP server on cloudflair. This is a direct result of SSE.

Loads of people have it but it is still rolling out.

Forcing every implementation to carry the complexity burden of the rare times you need streaming or notifications is not really a good tradeoff. If that is really required, it should be an optional extension to the implementation, not the core interaction.

Also, polling and paging are not the horror people think they are. It is less resource impactful than long lived stateful connections.

Also... I know I'm talking to an AI right now 🙄

The model response could be streamed. But the response from the MCP server doesn't need to be.

I think SSE is the problem here. Great they are moving away from that. 

where async bi-directional communication is necessary.

They should have left this out. It adds a ton of complexity and security issues with little gained. 

The complexity is in the transport. Until very recently that complexity meant you couldn't run it serverless for example. 

It is odd they didn't start with a basic http rest interface. Why reach for SSE when the ubiquity of rest exists?

There is a lot to like in MCP. Mostly I'm concerned about the transport. REST seems so obvious but it wasn't what they went for. That seems a little crazy to me given the ubiquity of REST servers.

I know nothing about the people building this. They might have a whole bunch of internal requirements. They might come at it from a perspective totally different than mine. I simply don't know. 

But I do get the feeling it has grown into something they didn't predict.

That is a huge amount of complexity to take on for that feature. It's not at all clear to me that server-driven events are worth it. In fact server-driven events kind of terrify me from a security perspective. There's a lot of benefit to not allow them and instead use something the client can control like polling.

Yeah, it seems crazy to me this wasn't the default from the start. It makes me worry about the maintainers honestly..

Sure but MCP clients only really support SSE or stdio. Libraries only support this as well.

I don't see the benefit of having the complexity of SSE as default instead of REST. The complexity trade off doesn't makes sense to me.

AI Studio is free. $20 for equivalent models is crazy. I know a few people who have, at least temporarily, paused their openAI sub to just use 2.5. 

I have no horse in this race and I certainly don't want to defend a giant company... But like... Inflation does exist... Seems there are a lot of people only now just discovering this.