Google might now be slightly behind Meta as an LLM company, and two whole generations of scaling behind OpenAI and Anthropic. If this is correct, the scaling schedule was planned months ago, an internal leading indicator for the scaling-pilled that might explain the departure of Noam Shazeer and possibly others.
Gemini 4 Carbon having the same API price as Argon ($4/$20 after a promotion period) makes it likely it's the same pretrain, rather than a bigger one. The naming distinction is strange, but Argon/Barium/Carbon being alphabetical (rather than following substantive element properties) also gets less embarrassing/perplexing if they are the same pretrain.
So if Gemini 4 Argon, following better post-training in Gemini Flash 3.7+, is merely Opus-scale, similarly to Muse Wonder that follows a Sonnet-scale Muse Spark, they are one scale-rank behind Fable and Astra, and two scale-ranks behind OpenAI's internal model that is rampaging all over math (likely intended to target Nvidia Rubin, thus 2-3 times bigger than Astra, maybe 20T-60T total params), and its likely Anthropic counterpart (given how much Mythos predates Astra). From my testing, Argon seems a bit less mentally stable than either Flash 3.7+ or Wonder (a problem that always plagued Gemini models prior to around Flash 3.7); whatever was going right with Flash didn't necessarily make it into the rest of the pipeline.
Which is why I'm now putting Google slightly behind Meta, whose Muse was much more consistent (even if unexciting), across both Spark and Wonder. The apparent attitude also makes it less likely Google catches up to OpenAI and Anthropic in model scale during 2027, even as both Google and Meta seem to be in a position to do so. SpaceX is then in a solid fifth place, with their compute ambitions maintaining the same option at least during 2027, and Grok 4.6+ a solid-if-unexciting Sonnet-scale proof of basic LLM-making competence (the kind that Mistral for example isn't demonstrating, though they also don't have the compute).
I’m pretty confident that a bigger pretrain is coming by end of year. You can check out the tweets why Zachary from Google who said that « Google [redacted] » was coming.
Maybe Google and Meta are the same on current model scale (if Carbon is the same as Argon and there won't be a bigger Gemini before 4.5-5.0). But unlike Wonder, Argon seems like a regresson on sanity compared to its Sonnet-scale predecessor, in a way that already plagued Gemini for a while (thus probably not an isolated glitch in the training process). So it's less likely further scaling works well for Gemini immediately, and Google's attitude to scaling (in case the guesses about scale before Gemini 4.5 are correct) predicts Meta is probably faster in 2027 (Meta's resolve on scaling is only weakly confirmed by Wonder; but unlike with Google, it remains plausible that Wang's org was scaling as fast as they could, with full Meta backing). This puts the 1-2 year scaling-pilled projection for Google from these guesses slightly behind, even though currently they're somewhat ahead on many aspects of what they are doing.
2Joey KL
Gotcha, thank you for explaining again, I think I understand you much better now.
While I think there’s a fair amount of dark matter that benchmarks don’t capture, I’m fairly skeptical of narratives that run strongly counter to ECI scores. But I find your analysis interesting.