I’ve had AI subscriptions running for well over a year now. Not in a reckless “sign up for everything” way, but in a deliberate, genuinely curious way — testing what’s actually useful versus what’s impressive in a demo and forgettable two weeks later. The stack I use today looks nothing like the one I started with, and I think that journey is worth sharing, because the tools I dropped weren’t bad. They were just beaten by better ones.
Here’s my honest account of where I’ve been, what I use right now, and why the most significant shift in my AI workflow happened not by switching services, but by moving off the consumer interfaces entirely.
How It Started: Gemini and Copilot
Through most of 2025 and into early 2026, I was running a split setup. Google Gemini handled images. Microsoft Copilot handled text. It worked reasonably well — Gemini’s image generation was solid at the time, and Copilot’s integration with the Microsoft tools I already used made it a natural fit for writing and summarising. Not exciting, but functional.
The problem with both of them was the same: the answers felt a bit safe. A bit corporate. Like they’d been trained to never say anything that might raise an eyebrow. For someone writing tech content, that kind of blandness is a problem. You need opinions. You need personality. You need the AI to engage with you rather than recite at you.
The Grok Phase
That’s when I started using Grok more seriously, and it immediately felt different. The answers were sharper. More direct. There was actual personality in there. If you’ve never tried Grok’s unhinged mode, do it once just to see what an AI looks like when it stops being polite — it’s genuinely entertaining, and it tells you a lot about how much personality most AI tools are deliberately suppressing.
Beyond the tone, Grok’s image generation was noticeably better than what I’d been getting from Gemini at the time. The “Imagine” feature produced more detailed, more interesting results with less prompt engineering. For a period, Grok was my go-to for both creative image work and anything where I needed a straight answer without padding.
The Switch to Claude and DALL-E 3
March 2026 was when things shifted again. I moved to Claude and DALL-E 3 almost full time, and the quality jump in the writing side was immediately obvious.
Claude’s approach to text is different from the others. It’s thorough without being watery. It holds a consistent voice across a long article, reasons its way through complex questions properly, and doesn’t hedge everything into meaninglessness. I use Claude Opus for the harder work — research-heavy articles, anything requiring careful structure, complex comparisons. Claude Sonnet handles the lighter stuff, the quick drafts, the rewrites, the ideas. Using both depending on the task makes more sense than using one model for everything.
DALL-E 3 was a significant step up from where I’d been for images. The prompt interpretation was better, and the results required far less iteration to get something usable.
Where I Am Now: The API Switch
April 2026 is where the biggest change happened, and it’s the one I’d least have predicted a year ago. I’ve moved almost entirely off the consumer interfaces and onto API-based calls. Instead of opening a chat window, the models run as part of my actual workflow — powering article generation, image creation, and content processing behind the scenes.
This includes switching from DALL-E 3 to GPT-Image-2 for image generation, and the difference is significant. The output quality is genuinely phenomenal. Sharper, more realistic, more responsive to complex prompts. It’s the best image generation I’ve used.
I should be honest about one thing here: working at the API level is more flexible and considerably more powerful, but it’s not a consumer experience. You’re writing code, structuring prompts programmatically, and managing outputs yourself. If you’re comfortable with a bit of technical setup, the freedom you get is worth it. If you’re not, the consumer interfaces are still very capable and there’s no shame in staying there. I run both — the API for the automated pipeline that powers this site, and the standard Claude interface for day-to-day writing and research.
My Current AI Stack
| Tool | How I Use It | Verdict |
|---|---|---|
| Claude Opus (API) | Complex articles, deep research, structured writing | Essential. Best for anything that needs real thought. |
| Claude Sonnet (API) | Drafts, rewrites, quick tasks | Essential. Fast and capable for lighter work. |
| GPT-Image-2 (API) | All image generation for the site | Phenomenal. Best image output I’ve used. |
| Grok | Checking in on new features, real-time X content | Still interesting, especially for tone and personality. |
| Microsoft Copilot | Dropped | Decent at the time, but outclassed. |
| Google Gemini | Dropped | Was useful for images early on; superseded. |
Total monthly spend on AI is considerably lower than it was when I was running multiple consumer subscriptions, because API costs scale with actual usage rather than charging a flat fee whether you use it or not. For anyone producing a consistent volume of content, that matters.
What Actually Lasts
The pattern I’ve seen over this past year is pretty consistent. The tools that stick are the ones that get out of your way and do the thing well. The ones that get dropped are usually the ones that prioritise the experience of feeling clever over the reality of being useful.
Claude has stayed because it’s genuinely good at what I need it to do, and the quality difference between Opus and the others is real for hard tasks. GPT-Image-2 has stayed because nothing else I’ve tried comes close on output quality right now. Grok is still interesting because it doesn’t take itself too seriously, which in this industry is rarer than it should be.
Hype Cycle Check
GENUINELY USEFUL NOW: Claude’s reasoning capabilities, particularly on Opus, have crossed a threshold where the output requires minimal editing for professional use. GPT-Image-2 for image generation is at a similar point. These are not experimental tools anymore.
WATCH CLOSELY: Local AI is developing faster than most people realise. Running models through something like Ollama on a reasonably powerful home machine is already practical for certain tasks. By the time CES 2027 comes around, I’d expect consumer hardware specifically aimed at on-device AI to be a major talking point. Local models mean no subscription, no cloud dependency, and better privacy. I’ve been testing this at home and it’s closer to daily-use quality than most headlines suggest.
APPROACH WITH CAUTION: Single-purpose AI tools with a monthly subscription and a narrow use case. Most of them do one thing that a general-purpose model can already do, and most of them won’t exist in 18 months. Unless the specialisation is truly deep and the use case is something you hit every day, the big platforms will absorb whatever niche they’re filling.
What This Means for CES 2027
Having been to CES around eight times, I’d say the show floor usually reflects where the industry was 12 to 18 months ago, not where it’s going. That said, I expect CES 2027 to be heavy on AI hardware — specifically devices that run models locally rather than depending on cloud calls. The interesting stories won’t be new chatbot announcements. They’ll be about what runs without an internet connection, at what speed, and at what price. For families thinking about privacy and ongoing subscription costs, that shift will matter. I’ll be there to find out which of it is real and which is CES theatre.
If you want to follow along as I keep experimenting with what works and what doesn’t, the Tech Dads Life newsletter goes out every Thursday. No hype, no sponsored content, just honest observations from someone who’s actually using this stuff. Sign up at https://techdadslife.beehiiv.com/ .
Which AI Tool Is Actually Right for You?
After a year of testing this stuff, here is the honest breakdown by use case — because the right answer genuinely depends on what you are trying to do.
If you just want one good general-purpose tool: Start with Claude on the standard consumer plan. The reasoning quality is the best I have used for text-heavy tasks, and the interface is clean and genuinely pleasant to use. You do not need the API or any technical setup.
If you write a lot and need consistent quality: Claude Opus is the version that earns the premium. For anything that requires genuine nuance — long articles, careful comparisons, structured research — the quality difference between Opus and the mid-tier models is real, not marketing. I use it for anything I care about getting right.
If image generation matters: GPT-Image-2 via the API is the best output I have seen, but it requires technical setup. If you want something consumer-friendly and still good, Midjourney remains strong. DALL-E 3 via ChatGPT is capable and easy to access.
For families and general household use: An Alexa or Google device with a general AI assistant is still the most practical entry point — nothing to download, no subscription to manage, and it answers the questions most households actually ask. The sophistication gap between a smart speaker and a dedicated AI tool only matters once you have specific tasks in mind.
For content creators and bloggers: The API route is worth learning even if it takes a few hours to set up. The combination of Claude for text and GPT-Image-2 for images covers almost every content production need, and API pricing is significantly cheaper than multiple consumer subscriptions if you produce consistently.
Things I would tell myself a year ago: Do not sign up for five tools at once. Pick one, use it properly for a month, then evaluate whether a second tool fills a genuine gap. Most overlap more than they differentiate. The tools that survived in my stack are the ones that got out of my way and did the job — not the ones with the most impressive demo.
The AI market is still moving fast. What I have written above reflects April 2026. Check back in six months and some of these answers will have shifted. That is not a flaw in the technology — it is the honest reality of a field still finding its footing.
