I went down a rabbit hole this morning reading the late-2025 Juejin picking roundups alongside the Google surface recap and the 2025年AI工具定价指南, and what crystallized for me is that almost every 2026 list I open is mixing at least three different billing models into one ranked column without disclosing which one the row sorts on. The picking roundups put Cursor at 综合能力第一 with a 20元/月 price and Codeium at 性价比之王 at 免费 as if those numbers are the same axis, the price guide bundles Gemini Pro at 19.99 with ChatGPT Plus at 20 and Claude Pro at 20 in one table, and the October trending recap calls Agent-S, claude-cookbooks, supermemory, and TradingAgents-CN simply 热门开源项目 even though they only meet inside a BYOK meter the rest of the list never asks about. How a tool is paid for decides what the reader owns when the session ends, and the roundups have stopped disclosing it.
The Google 2025年度最热门AI应用 recap describes how Gemini 3 has become five products — Gemini in Search, Gemini in Sheets, the Gemini app, Gemini in AI Studio, and Veo — and how the same account charges through any of those surfaces depending on the trigger. The December coding piece ranks Cursor at S档, Claude Code and Codex at A档, then folds Codeium, Tabnine, CodeWhisperer, and Blackbox into lower tiers on a five-axis decimal score without saying whether the rows include their own API meter. To be fair I would take the exact decimals and dollar amounts with a grain of salt because vendors have shuffled bundles since. The reading contract on the left says ranked tools, while the actual contract on the right is ranked billing models that happen to wrap a tool.
The meta-pattern I want to call out is that the roundups have been substituting a subscription tier column for a service-model column, and any honest comparison has to recover that distinction before the rank means anything. A 20 dollar Claude Pro subscription is a web surface and a prompt budget with one billing relationship with Anthropic; a Claude Code session on a BYOK router such as Claude Code Router charges against the user-provisioned key. Cursor at 20 per month pays for the editor loop and a fixed number of premium requests; Cursor Business at roughly 200 per seat swaps the meter for a centralized seat. Gemini Pro at 19.99 per month includes 2TB of Drive and a daily Pro budget, while Gemini API through AI Studio is a separate meter. Midjourney at 10 to 120 per month and Stable Diffusion run locally are obviously different contracts but the roundups print them on the same ranked list. My gut says any 2026 ranking that omits a service-model column is ranking the same word AI across at least four orthogonal meters.
Honestly I am a little skeptical of any 2026 roundup that prints one tier letter across Cursor, Claude Code, Codex, Codeium, GitHub Copilot, v0, Lovable, Gemini, ChatGPT, NotebookLM, and Agent-S without asking which meter the reader wants to be on. The picking scorecard ranks CodeBuddy at 9.6 and Blackbox at 7.2 across five axes that never include billing. The price guide offers dollar rows and yuan rows side by side. The October trending recap names Agent-S and claude-cookbooks without any API cost forecast. When an engineer tries to compose a stack they have to multiply 20 plus 20 plus 20 plus 10 plus the API spend nobody forecasted. That calculation is in none of the roundups, and I think it is the next layer down from the picking-tier-letter compression I have been writing about all month.
The practical takeaway I want to put down before the next quarterly roundup lands is that the picking list is still useful as an anchor, but the cross-tool decision has to start with the service-model matrix instead of the tier letter. The picking roundups are good for naming the editor-loop contenders (Cursor, Claude Code, Codex, GitHub Copilot, Codeium), the component-tier contenders (v0 and Lovable), and the mobile-app generators (Rork, VibeCode App, Anything). They are not good for telling me which vendor charges per seat, which charges per request, which hands the user a key to a different router, and which includes the AI in a non-AI bundle like Drive. Before I compare two tools I write down the meter — fixed sub, metered sub, BYOK, free, or local — because two tools at the same tier letter can have wildly different bills.
I will reassess in three months. For now I am still mostly on Cursor Pro plus Claude Pro plus an Anthropic API key for Claude Code sessions that exceed the Pro budget, plus Gemini AI Pro for the Drive and Sheets features I already use. What has changed is that I read every roundup as a billing list dressed as a tool list, and when the post skips the meter I treat the tier letter as marketing. Give it six months and I expect the better roundups to add an explicit meter column — fixed sub, metered sub, BYOK, free local, bundled enterprise — and when that lands I will trust the rank across rows. Until then I treat the scorecard as a contact list of vendors and the meter as the engineering decision, and I will reassess in three months whether the format has caught up.
Top comments (0)