In my experience, Google AI has finally caught up with the frontier, plus it uses Google search for facts. Now I often find myself used up my Gemini/antigravity quota quicker than Codex/chatGPT. Of course Opus 5.5 is still best, but Argon got its edge.

I have suspected this widespread waste since Strata is out, But it is good to have the suspicion confirmed by Niko. Tens times cost reduction/ or token production expansion imminent, depending on the lenses one uses.

Put in perspective, a few years ago, 30+% growth in cloud, which all what the market cares back then, was considered "HIGH GROWTH" Now, we see business doubling every year, literally sellers cannot keep up with demand. "DOOM, BRACE FOR IMPACT".

Fast, cheap, good, cannot get all three of them, right? Strata now runs Qwen 3.8 Flash on a gaming PC fast enough that people ask the obvious follow-up: is it actually good? The honest answer is that we run 3-bit packs (IQ3_S, IQ3_XXS), so there is a gap to the official API. We measured it. The official Qwen 3.8 Flash (Alibaba's API) answered 31 prompts, about 15K tokens, and Strata read the same answers token by token: Ask the official API the same thing twice at temperature 0, and it disagrees with itself on 2.5% of tokens. That's the ceiling. Local adds 1.5 points on top of that with IQ3_S, 2.3 with IQ3_XXS, and 2.6 with our fastest setting. Where the official model is sure (about 70% of tokens), local picks the same token 99.9% of the time. The rest is mostly synonym-level choices, where the API wavers too. And the pelican riding a bicycle? No sweat. The real perk of local: you know exactly which model and which weights you're running. No asking for Astra and getting quietly routed to Luna. No model shrinkflation. Fast, cheap, and good, we take at least 2.5 of them.