I went down another rabbit hole this morning, this time reading the late-2025 and early-2026 AI product roundups on Juejin, and one pattern jumped out harder than the others. The "which AI tool is best" lists are starting to blur together because the products themselves are converging, and the things actually worth arguing about have moved upstream. I want to write that down before I forget, because every time I read one of these roundups I end up talking about the wrong question.
The first thing that stuck me is how similar the top of every list looks. GitHub Copilot, Cursor, Claude Code, 通义灵码 from Alibaba, Google's Antigravity, Codeium — they all show up in roughly the same order, with roughly the same feature checklists, and roughly the same pricing tiers. The Cursor numbers I keep seeing (something like 35% market share and 2M+ paid users, with a rumored 500M ARR) are loud, but the gap between it and the second tier is shrinking. Antigravity in particular is the one I want to actually try, because the pitch of "agent-first editor built on Gemini 3" is something I haven't felt from a Google dev tool in years, and to be fair Google has killed more beloved dev products than I can count, so I am a little skeptical. The fact that it keeps trending is doing real work, though, and I'll give it a serious run next month.
Claude Code keeps being the one people I trust actually use for the hard stuff. The planning step, the sub-agents, the way it walks you through a refactor of a 14-file auth module without losing the thread — that experience is genuinely different from anything in the IDE-native tools, and the terminal-first flow is starting to feel like the default for big changes rather than the niche. Cursor remains my daily driver for the boring 80% because the Tab key flow and the inline diff UI are still the fastest thing on the market. Copilot wins on the boring metric of "it is already there" — when your team lives in VS Code and your PRs live on github.com, the switching cost of anything else is real, and I think people underestimate how much of the AI tool conversation is just inertia.
The Chinese tools are the part I think Western engineers are underrating. 通义灵码 keeps showing up with a pitch that it matches GPT-4o at lower enterprise cost, and for any company already inside the Alibaba Cloud ecosystem the integration story is brutal to beat. DeepSeek-V3 keeps coming up as the model underneath half the new tools, and the inference cost math is hard to argue with if you are running anything at scale. I have not stress-tested these in production myself, so I'd take my read with a grain of salt, but the rate of improvement from 2025 to now has been faster than I expected and I don't see it slowing down.
The thing that actually changed my mind reading these lists is the workflow argument. A few posts argued, persuasively, that the next bottleneck is not the model or the IDE — it is the input you hand them. Mermaid diagrams for data flow, explicit API contracts, generated ER diagrams, a short spec for the change you are about to make. Documentation-driven development as a way to feed the agent, not as a relic of waterfall. Honestly I had been doing this for months without a name for it, and seeing it spelled out made the version I was doing badly feel fixable. To be fair, I am not fully sold on writing a spec for a two-file change, because most of my day is two-file changes. I think the right version of the argument is that for anything that touches more than three files or lasts longer than a week, the upfront spec pays for itself, and the tools that handle that input well are the ones I will keep using.
The meta-takeaway I keep landing on is that the "best tool" question is becoming less interesting than the "what workflow" question. The tools are converging fast. Most of them can do most of the things most of the time. What actually differentiates your experience is whether you can structure the problem, whether you can review what the model produced, and whether you understand the code underneath. None of these tools save you from that, and the people I see getting the most out of them are the ones who were already decent engineers.
I'll reassess in three months. That's the only honest timeline I can commit to given how fast this space is moving, and the last time I said that I was bouncing between Cursor and Claude Code, which is still roughly where I land today. If you're picking something right now, my advice is the same as it was six months ago: don't overthink it, pick whichever one you'll actually open every day, and save your energy for the parts of the job the model can't do yet.
Top comments (0)