On new models achieving AGI

OpenAI released GPT-6 Astra last week only days after Anthropic released Fable/Mythos 5.1.

Anthropic was showing how their models were the most powerful yet only to be trumped by OpenAI 2(!) days later. Many started claiming that Astra is now AGI as it got almost a perfect score on ARC-AGI-3.

I don't think I can agree with that.

As can be seen in comment threads all over the web the term AGI doesn't have a clearly defined definition.
Is it supposed to excel or simply suffice in task solving?
Should it solve novel problems?
Are only computer tasks included in the skill set?
Should it be able to reason about complex problems and produce new insights?
Should it be aware?

I can't answer those questions, I leave that to people more intelligent than me but what I can analyze is how the models are benchmarked and presented.

Usually when it comes to presentation on performance, design or efficiency of a product I tend to not put too much weight on what the marketing material. Of course it will show how much better product X is than Y. It is marketing after all.
If any your product is measuring worse then the competition on a metric you don't include that metric in the release announcement, it would be honest to do so but when the amount of money at stake is as high as it is for the AI industry you'll play it safe.

Just now Artificial Analysis updated their numbers on the new models and it scores Fable 5.1 higher than Astra which seems odd if Astra is the first model to achieve AGI. Why would Anthropic not push on their score on ARC-AGI-3 if their model is as intelligent or maybe even more intelligent (according to this measurement) than Astra?

I think the answer is the usual one when it comes to benchmarking and marketing, they didn't score as high on that particular benchmark. Then how come a more intelligent model isn't scoring as high?
It is probably down to benchmark maximizing, OpenAI has probably trained their model specifically for that benchmark while Anthropic hasn't. This isn't something new to the world of benchmarks, the Volkswagen scandal years ago comes into mind.

What actually matters is the performance users experience, real users - not tech influencers that got early access as they tend to be part of the hype train.

I don't have access to Astra and it would be unfair of me to do a review on it. What I can do is give my opinion based on my work with other models and as recently as a few weeks ago Opus made the hilarious mistake of naming a folder * in my repo, it realized the mistake and decided to delete it. Fortunately my sandbox had it ask for permission to run rm -rf *. I would expect a mistake like that from a *nix beginner but from a one of the most expensive frontier models?

I don't think we have AGI just yet.