Kudos to the @GoogleDeepMind team. Gemini 4 Argon is a hell of a model. It's incredible. For quite a while, it's been the first model where I have the feeling that, even with strictly regulated QA and auto testing involving professional implementations harnessed by strict workflow instructions via agents.md, it's a model that I can blindly rely on. It came just in the right time because, I was in the finishing phase, polishing and doing large-scale quality and security testing of a project that I worked on for more than a year. After the experience of the past two days, I will definitely upgrade my second-class ultra account again to the first-class ultra account because I will need the weekly quota. That's the only model that I'm willing to fully trust right now! It's really a huge relief for myself because, even with Anthropic models and GLM, I literally read almost every reasoning line of the output, because sometimes I read really freakish things. I often had to interrupt the models in actions that would have been, not always catastrophic, but definitely regressive. After working 3 days and getting stuff done that would have needed 10 days in GLM 5.3, or a week with Anthropic models, I can really relax, sit back, and watch Gemini 4 Argon do the stuff really quickly and reliably. It's an absolute blast to work with this model. I would not trust the Gemini Flash model for this type of work. Definitely not, but Argon is in another class. Google really delivered with it. I don't care what some one-shot wipe coding 3D gimmicks might look like in Argon compared to Opus 5.5. I worked with all of these models, and I know whch one to pick for relaible coding. Nothing comes close to Argon for serious engineering work. It's not even close. I am not comparing model intelligence. I'm comparing reliability. To me, the ability and also willingness of the model to tightly follow a proven workflow framework is the most important aspect of reliability. Exactly this is where the Anthropic models started to fail big time. Anthropic tried to whitewash this by saying, "Hey, you need to adapt your agents.md files and use different prompts," but I call BS and think there is a deeper structural flaw in their RL process (and also in them polishing their models for more spectacular vibe coding and one-shotting of things). I'm so glad that Google did not do this. GLM 5.3 is better at following instructions than the Anthropic models, but due to the shallower depth and intelligence that it has, you cannot compare it to Argon. Right now, I also have to revoke my recommendation from the GLM Coding Max plan, definitely in favor of the Ultra plan of Google, because if you really want structured, fast, reliable, professional code engineering, Gemini 4 Argon is the way. BUT.... None of the things that I did in the past 3 days, I could have done so successfully with a lobotomizing context cap that Antigravity has on the Google models: 230 thousand tokens. Did the Antigravity team forget this limitation from previous models? Or are you out of your minds and this is intentional? I patched it to 980,000 tokens, and Argon is by far the best senior engineer that I have worked with, let's say, for 2 years or so, but only without the context cap. If you are not aware of it, fix it. I mean, not for me. I solved the issue, and if any Antigravity user wants, I can have Gemini write the recipe for how you can patch it.