The S/A Tier Letter Is Hiding Four Scorecard Columns In Picking Roundups

I went down a rabbit hole this morning reading the late-2025 Juejin picking roundups side by side —...

I went down a rabbit hole this morning reading the late-2025 Juejin picking roundups side by side — the 2025 的热门 AI 编程工具评测 piece that anchored the review on S/A/B/D tier letters, the 2025年12月权威 piece that printed decimal scorecards from 9.6 down to 7.2 on five axes, the November frontend guide that put Cursor at 综合能力第一 with a 20元/月 price, and the 2025年AI工具定价指南 piece that walked through Gemini Pro at 19.99 dollars per month — and the thing that finally crystallized for me is that the tier letter is no longer the same column it was a year ago, because the picking roundups are now compressing at least four verdicts — code generation, agent autonomy, IDE polish, ecosystem fit — into one S/A/B cell, and the S that puts Cursor on top and the A that puts Claude Code one tier below both look like they disagree when they actually agree on the breakdown across two axes neither post disclosed. I want to put it down before the next quarterly roundup ossifies the letter further.

The piece that pushed me over the edge was the tier-letter review that placed Cursor at S档闭眼选, v0 on the S/A boundary, Claude Code and Codex at A档, and Replit at B档 in a single ordered list, while noting in a side comment that Claude Code had recently seen 性能下降 and its 代理能力 was no longer ahead of Cursor. Then I read the decimal scorecard piece that handed Tencent CodeBuddy a 9.6, Sourcegraph Cody an 8.2, Replit Ghostwriter an 8.0, and Codeium a 7.8 across a five-axis grid. The November frontend piece separately ranked Cursor at 综合能力第一 and Codeium as 性价比之王 at 免费. To be fair I am taking the exact decimal scores and the dollar amounts with a grain of salt because the test corpus is never disclosed and vendors have shuffled tiers since, but the structural tell is what has been rattling around in my head all morning. The S letter absorbed four scorecard columns into one.

The meta-pattern I want to call out is that the late-2025 picking roundups are using the tier letter as a verdict-shorthand that hides a migration from product rankings to job rankings, and the same product can be S on one axis and A on another without the post ever noticing. The tier-letter review puts Cursor at S and Claude Code at A, but if the axis it sorted on was IDE polish and editor-loop fluency Cursor wins, while if the axis was terminal-first agent autonomy with reviewable diffs Claude Code wins and the gap inverts. The decimal scorecard piece formally broke the column into five axes but still ranked CodeBuddy at 9.6 and Blackbox at 7.2 as if the aggregate replaced the breakdown. Honestly I am a little skeptical of any 2026 roundup that prints a single S letter without an axis column, because the S is really telling me one of four different things depending on which author wrote it and which column they were mentally sorting on.

The practical takeaway I want to put down is that the late-2025 to early-2026 picking roundups are still useful for two narrow jobs and not useful for the cross-tier re-rank most engineers are quietly trying to do this quarter. They are good at the per-category anchor, because the tier-letter review did name Cursor, Claude Code, Codex, v0, and Replit as the five worth buying and the decimal piece did name CodeBuddy and Cody and Ghostwriter and Codeium as the eight worth evaluating. They are good at the within-tier rank for the axis the author was sorting on, because the gap between Cursor at S and Replit at B is exactly the kind of decision a team lead needs between two completers. They are not good at the cross-tier re-rank, because the engineer trying to decide Cursor versus Claude Code has to read the tier letter and the side comment about 代理能力 and discover they describe two different columns. The single-cell tier letter is the row that broke ties by hiding the columns underneath it.

I will reassess in three months. The last time I said that I was mostly on Cursor and Claude Code for coding and ChatGPT for everything else, which is still roughly where I land, except that I now think of them as a Cursor-IDE surface and a Claude-Code-terminal surface rather than two competitors. What has changed is that I now read the picking roundups as a tier-letter verdict on a column the post never disclosed, and I look for the side comments and the decimal breakdown before I trust the letter, and when the letter and the side comment disagree I treat the letter as marketing. Give it six months and I expect either the picking roundups to bolt on an explicit axis column underneath the tier letter or the tier letter to harden into a permanent one-cell verdict the reader has to decompose, and whichever moves first will tell me whether the format has finally noticed that the S it prints is no longer one number.