“What’s the best AI?”
I get this question constantly. From clients, friends, other builders. And the honest answer is annoying: it depends on what you’re doing.
There is no single AI that’s best at everything. Not yet. Maybe not ever. The landscape right now is fragmented, fast-moving, and genuinely competitive. Different tools win in different categories.
Here’s what I’m actually using in February 2026. Not theory. Not benchmarks. What’s in my daily workflow right now.
Coding
This is where I spend most of my time, so it’s where my opinions are strongest.
My primary setup is Claude Code with Opus 4.6. I run it in the terminal alongside my projects and it handles everything from scaffolding components to refactoring entire modules. The context window is massive. The reasoning is sharp. When I need to think through architecture or debug something gnarly, Opus doesn’t just give me code. It gives me understanding.
For my IDE, I use Cursor. Inside Cursor, I’ve been running Codex 5.3 and the newly released Composer 1.5. Codex is reliable for inline completions and quick edits. Composer 1.5 has been impressive for multi-file changes where you need the AI to understand the full picture of what you’re building.
The combo of Claude Code in the terminal and Cursor with Codex/Composer in the editor covers pretty much everything. Terminal for the heavy lifting. Editor for the flow state.
Writing
For writing, I’ve landed on Kimi 2.5. I run it locally and use it in the cloud.
This one surprises people. But Kimi’s writing output just feels different. It’s less robotic, less predictable, less “AI-sounding” than most models. When I need help drafting, expanding ideas, or working through how to frame something, Kimi gets closer to how I actually write than anything else I’ve tried.
I still write everything myself. The AI is a collaborator, not a ghostwriter. But having a writing partner that doesn’t sound like every other LLM is worth a lot.
Image Generation
I use AI image generation mostly for assets. Social content, blog headers, concept work for client projects.
Three tools in rotation:
Midjourney is still the one to beat for aesthetic quality. The images look intentional. They have style. When I need something that looks like a human with taste made it, Midjourney wins.
GPT-Image-1.5 is great for speed and flexibility. The prompting is more natural, and the results are consistent. When I need something good fast, this is where I go.
Nano Banana Pro has been a quiet favorite. It handles stylized content well and the outputs have a distinct look that works great for certain use cases. Worth exploring if you haven’t tried it.
Each one has strengths. I pick based on what the project needs.
Research and Thinking
Back to Opus 4.6 in Claude Code.
When I need to think through a problem, research a topic deeply, or reason through a complex decision, this is my go-to. Not a chatbot conversation. An actual thinking partner.
I’ll throw it a messy problem and it will break it apart, consider angles I missed, and push back when my reasoning has gaps. The depth of analysis is what separates it. Other models give you answers. Opus gives you thinking.
This is the tool I use when the stakes are highest. Strategic decisions, architectural choices, anything where getting it wrong costs real time or money.
Agent Autonomy
This is the frontier. AI agents that can operate with real independence. Not just answering questions but doing work.
Two models stand out here: Opus 4.6 and Codex 5.3.
Opus handles complex, multi-step autonomous tasks better than anything else I’ve used. It maintains context across long chains of reasoning and action. When I set it loose on a codebase with clear instructions, it delivers.
Codex 5.3 complements this well. It’s reliable for execution-heavy agent tasks where you need consistent, accurate code generation at scale.
The agent space is moving fast. What works today might not be the best option in three months. But right now, these two are carrying the load.
The Honest Take
There is no right answer.
I know that’s unsatisfying. People want a clean recommendation. “Use this one.” But that’s not how it works in practice.
The best AI for coding isn’t the best for writing. The best for image generation isn’t the best for research. And the best for any of these categories six months ago isn’t necessarily the best today.
What I’ve learned is this: the builders who get the most out of AI aren’t loyal to one tool. They’re fluent across several. They know what each one does well, where each one falls short, and when to reach for which.
That’s the real skill. Not finding the best AI. Knowing which AI is best for what you’re doing right now.
Build your own stack. Test things yourself. The benchmarks don’t tell the full story. Your workflow will.
What’s your current AI stack? Genuinely curious what’s working for others.