You need to use search grounding.

They need to pretrain on fresh data.

or, they've had 4 for a long time (which they did). AFAIK they have had 5 for a while now too.

I'm not quite sure what their strategy is, because it looks like the models released just have tweaks on the same original training data, but nothing more. Basically, they decide not to release them and it doesn't seem to be a much different when they finally to decide to release.

Seems like a rushed RL run on top of an old model to jump into the hype train, seems reasonable since they hold a fantastic distribution moat. Even if it's not SOTA a billion people will use it and they'll collect data-points for future iterations. Same as muse.

Knowledge cutoff isn't really what these models are used for.

How is a model that has never seen an agent on it's pretraining data going to reasonably plan and fan-out tasks to multiple agents?

I don't think that's likely to be a problem, really. If I explained the concept of an agent to you in a couple of sentences in Feb 2025, could you understand it, even though agents didn't exist yet? I think you could, and I think a modern frontier LLM probably can too.

Yeah, it can be worked around with in-context instructions but they still tend to scope, delegate and architect without awareness of the fact that agents exist and work. It's an actual big deal IMHO, opus 5.5 is the most updated AFAIK.

Nowadays the scraped training data is only one part of the training process. The RL phase of training can instill a knowledge of tactics like sub-agent delegation without documentation about them being present in the training corpus. Context isn't the only way for a model with an early knowledge cutoff to know about it.

Still a crutch around incomplete pre-training data, a ton of useful knowledge wont be properly represented on the weights.

Exactly my first test on all Gemini models. It cannot recollect any of the top events since Feb 2025. The worst about this is that since 3.6 Flash Google devs stated "knowledge cutoff March 2026" in the system prompt, so the model is wrongly confident that it knows about the year that it was not trained on. Which means that it searches less than it would.

And at least Flash 3.7/3.8 were both difficult to persuade that they have wrong knowledge cutoff.

I reported it as a bug to Google, but as all of my reporting to Google, no response ever. This is just stupid!

They know it and they try to cover it up with a fake reported cutoff date. Claude reports June '26 but it has only partial 2026 data (tested it against the API and it's 25 as well). GPT refuses to answer but it's old 25 data.