Google may be holding back Argon over compute limits, developer says
An analysis post argues model competition is really a compute-scaling game and speculates that Google is delaying Gemini 4 Argon because of compute constraints and its supply commitments to Anthropic.
It is all a compute scaling game. Claude was doing well with the 4.x series when they overbought compute for training. Then OpenAI's datacenters lit up, they were able to catch up with the 5.x series and give it out for a lot cheaper (higher limits) and ran faster. For most of the middle of 2026. And by August they pulled what was unimaginable, many companies spent less on Anthropic in Aug/Sep than they did in July (it is true, I have see many internal billing dashboards - part of it was Anthropic was getting too costly, and part was OpenAI as alternative existed) Then Anthropic bought compute from XAI, Google (even asked Meta, but they said no). And they launched the 5.x series, which was a step change in abilities, but more importantly, cheaper, faster and higher subscription limits! And the tide turned again. From Sept to Oct, the revenues moved from OpenAI to Anthropic again! Meanwhile Meta was building up compute like crazy in 2025 (I was there) and both internally and externally there were questions on whether it is a good idea after teh Llama 4 flop. Who will use it? What is the point? Will we just sell the compute. But that compute build up is what allowed training Muse (Spark, Wonder and the future models), and Muse Spark could be given out for very very little cost for very high limits again. And Muse could HARD-launch in the public without any signup limits. I suspect Google is unable to release Argon because they do not have compute, and they have commitments to Anthropic (safety is probably a hogwash argument, Opus, Wonder are safe enough at similar capability ceiling)