It is late July 2026 as I write this. A year ago I had several AI subscriptions running at once. Perplexity for research. A multi-model hub for “access to everything.” A couple of niche writing tools. I told myself this was rigor — using the best tool for each job. Mostly it was just tab-switching dressed up as strategy.
Here’s what that year actually taught me: the AI industry keeps selling a story where the endgame is one enormous, all-knowing black box that does your thinking for you. That story is wrong, or at least beside the point. What’s actually changed is something quieter and more useful — a single person can now move across disciplines that used to require a team. You don’t need to have memorized a citation style or a syntax spec to work like someone who has. The AI handles the narrow technical execution; you supply the judgment, the taste, the “does this actually make sense” instinct, because you are the one who has a soul inside.
But that only works if the tool gets out of your way. And that’s where most of my stack failed me.
The Problem With Having Everything
Multi-model hubs — the ones that give you a dropdown with GPT, DeepSeek, a dozen image generators, whatever’s trending — look great on paper. Every model, one subscription. I had them all with Abacus ChatLLM for only $10 a month. What’s not to like?
In practice, that dropdown is a tax you pay every single time you sit down to work. Before you can think about the actual problem, you have to think about which tool to use. That’s decision fatigue with extra steps, and over a year it adds up to a real drag on momentum. Or you can let the aggregator decide for you.
Big mistake. It always went for the most expensive model (Claude, at the moment), squandered all the credits on preliminaries, then fell back to free models, leaving me with projects where the end did not understand the beginning. Nothing I could tell it or force it to do via software sliders made it deviate from this behavior. In the end, I figured the free models were good enough with a lot of handholding and just used those. At least, I got consistent code.
There’s a deeper issue too: these hubs mostly give you raw API access, model to model, with none of the connective tissue. No real memory of your files. No sense of what you were doing yesterday, or five minutes ago. Every session starts from zero. What’s worse, the lack of any sort of project memory meant that every prompt required everything preceding it to be sent out again. Token consumption grew exponentially under the hood. No wonder the credits burned so quickly.
What Actually Stuck: Gemini
I didn’t set out to consolidate around Gemini. It just kept being the thing I reached for in order to be consistent with what I already had, until one day I noticed I’d quietly stopped opening everything else.
I stopped using Perplexity. Perplexity’s whole pitch was bolting search onto a chatbot. Gemini doesn’t need the bolt-on — it’s built retrieval-first, sitting right on top of Google’s index. Ask it something that needs current, sourced information, and it just goes and gets it, inline, without me having to leave and come back.
Nothing lives in a bubble anymore. This is the part people undersell. Gemini sits inside the same ecosystem as many of my documents, my drive, my actual working files. I’m not copy-pasting paragraphs between a chat window and a doc fifteen times a day.
Ah, and Gemini has a one-million-token context window. How about that?
About Those Million-Token Context Windows
Every AI company loves to brag about how much text you can cram into a single prompt now — like context size alone is a proxy for intelligence. It isn’t. Dump 500 pages of unsorted material into any model and you’ll watch it lose the thread somewhere around page 200, or start inventing details to fill the gaps. Bigger isn’t smarter. It’s just bigger. In the case of my code project, it began to forget the code that had been generated for certain features, and began producing new code stripping those features entirely. Much later, Gemini (Flash, coding was done with Pro) explained that this is expected behavior because old context corrupts new context unless the whole thing is tightly focused.
What actually works is curated context — a focused batch of the specific papers, files, or code that matter for the task in front of you. Gemini handles that kind of mid-sized, relevant cluster well. But the curating is still on me. The model brings speed; I bring the editorial judgment about what deserves to be in the room at all.
I Don’t Want an Autonomous Agent
There’s a real push right now toward AI “agents” that take over entire workflows — book my trips, run my project, manage my inbox, decide things on my behalf. I get why that’s appealing to a certain kind of org chart. It’s less appealing to me. I’m not ready to be turned into a moron that makes no decisions and spends all his time on social media.
I don’t want a black box making calls I can’t see. I want a very fast, very capable collaborator that does the grunt work — pulling research together, drafting the boilerplate, compiling the data — so I can spend my attention on the parts that actually require a human: judgment, synthesis, deciding what matters, beauty.
So, After a Year
I don’t need need twenty specialized tools. I need one that disappears into the background instead of demanding I manage it. Gemini Plus (at the lowest level that will give me decent AI limits, which costs $4.99 a month at the moment) replaced my research tool, my file-shuffling routine, and most of my decision fatigue with one calm, quiet interface — and left me doing the part of the work that was mine to do in the first place: the thinking. I admit that it helped me to write this article, starting from a chat where I questioned its value. I asked it to summarize it, insisted that certain things that were said should remain explicit, took the product and give it to Claude (free level) for humanizing, and then pasted the result onto my blog. A few more edits and I was done.
Are you having a similar experience with AI? Are you letting it tell you what you should do? Even if the AI is correct, it’s kind of disrespectful on its part, obsequious style aside.
But again, it doesn’t have a soul. It can only tell the difference between right and wrong, between beautiful and ugly, by running statistics.
You, on the contrary, know it from the inside. Because you do have a soul.
Update 8/26:
Barely a couple weeks since I wrote this, and I already have something to change: I’m hitting usage limits with Gemini Plus (the $4.99/month plan) which I didn’t hit with Gemini Pro. I asked Gemini itself, and it explained that because every prompt re-processes the entire chat history, heavy coding sessions quickly burn through the much smaller hourly compute quotas and tighter context limits of the Plus tier compared to Pro. I was not reaching the limit before simply because the Pro level has a massive capacity cushion. While competitors like Claude boast larger default windows that delay the problem, you still hit walls eventually. To get around this without losing my place, Gemini suggested a manual workflow: prompting it to ‘Reset Context’ to generate a clean summary so I can spin up a fresh chat with decent continuity.