# After Opus 5.5, Claude Code users still complain about the length or clarity of its replies at three times the field

*Anthropic's 22 September post said Opus 5.5 answered the complaint that Opus 5 was hard to follow. From 28 September to 8 November, 174 of 15,227 people who judged Claude Code still complained about the length or clarity of its replies, 3.0 times the other coding agents.*

Capability case, Claude Code. By Feedback Bench, 16 November 2026. Feedback Bench 1.1. Posts 31 August 2026 to 8 November 2026. Web: https://feedbackbench.com/coding-agents/#/articles/after-opus-5-5-claude-code-users-still-complain-about-the-length-or-clarity-of-its-replies-at-three-times-the-field

On 2 October the Claude Code team announced a plugin called You should Know, which scans Claude's output for important information you might miss. One of the replies came from a developer who already knew the problem:

> yes please. half my debugging is scrolling back for the one line that mattered while the agent kept talking.

[A user on X, 2026-10-03](https://x.com/idohodl/status/2106371608747167803)

That was ten days after Anthropic's 22 September post for Opus 5.5, which said early testers found its writing clearer and easier to follow, and that this addressed common feedback about Opus 5.

In the posts we read, that complaint did fall after 22 September, and by a lot. It did not go away. From 28 September to 8 November, people who used Claude Code still complained about the length or clarity of its replies at three times the rate of people using other coding agents.

## Before and after the release

In the three weeks before the release, 404 of the 9,133 people who judged Claude Code from their own use complained about its replies. That is 4.4%. Among people who judged other coding agents, it was 0.35%.

The week of the announcement was already lower. In the week of 21 September, 90 of 4,797 people complained, or 1.9%.

From 28 September through 8 November, 174 of 15,227 people complained. That is 1.14%, about a quarter of the rate before the release. Over the same weeks the other coding agents stayed close to where they were, at 0.38%.

**Exhibit 1: After Opus 5.5, Claude Code's rate fell but stayed three times the field.** Share of people judging each side who complained about the length or clarity of replies, in the three weeks before the 22 Sep Opus 5.5 release and from 28 Sep. The release week of 21 Sep is left out. Other coding agents are pooled, with Claude Code's judges removed.

|  | Share of people judging |
|---|---|
| Claude Code, before 21 Sep | 4.42% (404 of 9,133) |
| Other coding agents, before 21 Sep | 0.35% (71 of 20,357) |
| Claude Code, from 28 Sep | 1.14% (174 of 15,227) |
| Other coding agents, from 28 Sep | 0.38% (102 of 27,037) |

*31 Aug to 8 Nov 2026. One person counts once per side and period.*
The same complaint rate before 21 September (404/9,133 against 71/20,357) and from 28 September (174/15,227 against 102/27,037), Claude Code against the pooled field. The release week is left out.

The labels don't say which model anyone was using, so these counts can't say what caused the fall. Opus 5.5 came out on 22 September, Sonnet 5.5 on 28 September, and the Concise output style a month earlier. What they do show is the timing: the fall starts in the week of the Opus 5.5 post.

## The gap that is left

After the release, 1.14% against 0.38% is still 3.0 times. That holds against the larger agents one at a time. Codex users complained at 0.47%, so Claude Code is 2.4 times Codex. OpenCode is at 0.43%, and Claude Code is 2.6 times that.

**Exhibit 2: From 28 Sep, Claude Code is above Codex, OpenCode and the field.** Share of people judging each agent who complained about the length or clarity of replies, 28 Sep to 8 Nov. Anyone who judged Claude Code in the window is removed from the other rows. Other coding agents are pooled.

|  | Share of people judging |
|---|---|
| Claude Code | 1.14% (174 of 15,227) |
| OpenAI Codex | 0.47% (53 of 11,276) |
| OpenCode | 0.43% (21 of 4,841) |
| Other coding agents | 0.38% (102 of 27,037) |

*31 Aug to 8 Nov 2026. One person counts once per row.*
Complaint rate from 28 September: Claude Code 174/15,227 (1.14%), Codex 53/11,276 (0.47%), OpenCode 21/4,841 (0.43%), field 102/27,037 (0.38%). Claude Code's judges are removed from every other row.

This does not make Claude Code the highest on this complaint. Kiro had 2 complainers among 245 judges, too few to tell apart from Claude Code's rate.

Most of the complaints came from Reddit, 145 of the 174. On Reddit alone, Claude Code is 2.3 times the other agents. Leaving Reddit out, it is 4.7 times. On X it is 3.7 times.

Praise on the same point is rarer. In the later period 49 of the 15,227 people praised the length or clarity of Claude Code's replies, and two of them also complained.

Every week from 28 September sits between 2.4 and 5.4 times the other agents, so this is not one bad week. The most recent week, starting 2 November, is the highest since the release: 43 of 2,907 people complained, 1.48%, which is 5.4 times the field.

## What the docs already say

Claude Code already has a Concise output style. It shipped in the week of 17 August, before these posts, and the docs say to use it when default responses are longer than you want.

The Opus 5.5 post names the complaint as something users said about Opus 5. Neither page says how many people make it, or how that compares with other coding agents.

## How we counted

The build is labels on `main`, weeks of 31 August through 2 November 2026, so posts run through 8 November. Counts are people, not posts. One author string is one person, and every judging post that person wrote about an agent folds into one yes or no.

A person judges an agent when a post is marked as judging, is about that agent, passes the firsthand check, and the author is not dropped by the author screen (automated for every agent; staff, promotion or declined for that agent) or by a staff handle on the agent.

A complaint is a `work.response_verbosity` label with polarity complaint. The criterion is "Length and clarity of replies, summaries and comments." It covers a reply that was too long, too short or hard to read, and the label does not separate those.

The periods come from the release date. Before is the weeks of 31 August, 7 September and 14 September. The release week is 21 September, which contains the 22 September post. Later is the weeks of 28 September through 2 November. Across the whole window the rate is 6.0 times the field, but that mixes the weeks before and after the release, so we do not use it as the headline.

The field is every person who judged another ranked coding agent in the same period, with anyone who judged Claude Code anywhere in the window removed, so the two sides share no one. Each comparison with a single agent removes the same people. The test is Fisher's exact test, two-sided, on people. The later-period gap has p = 8.2×10⁻²⁰. The fall from 4.42% to 1.14% has p = 2.2×10⁻⁵⁷. The corrected bar for the eight candidates we scored is 0.05/8 = 0.00625, and the scan-wide bar is 0.05/540 = 9.3×10⁻⁵. Codex (p = 1.8×10⁻⁹) and OpenCode (p = 3.0×10⁻⁶) clear both. Kiro's comparison is p = 1.0.

The busiest author in the later period wrote 4 of the 189 complaint posts. The week of 12 October is the weakest later week, at 2.5 times (p = 0.0068). That misses both the candidate bar and the scan-wide bar. The claim rests on the six-week pool. The X-only cut (p = 2.6×10⁻⁵) clears both bars. The Reddit-only and without-Reddit cuts clear both. The label file has no thread id or subreddit, so dropping every Reddit author at once is the check that no single subreddit carries the result.

The opening quote is from a post marked as a clear complaint in the committed quote check (`data/labels/quote-checks-2026-08-31.jsonl.gz`), read on X on 16 November 2026.

The docs pages were read on 16 November 2026. Neither states a complaint rate.

## Sources

- Feedback Bench coding labels, `data/labels/labels-2026-08-31.jsonl.gz`, firsthand file, and author screen, weeks of 31 August to 2 November 2026. Recount: `recount.py` in this folder.
- Criterion `work.response_verbosity`, "Length and clarity of replies, summaries and comments", in `pipeline/criteria_v1.json`.
- Anthropic, [Claude Opus 5.5](https://www.anthropic.com/claude-opus-5-5), 22 September 2026, read 16 November 2026.
- Claude Code docs, [Output styles](https://code.claude.com/docs/en/output-styles), and [What's new, week of 17 August 2026](https://code.claude.com/docs/en/whats-new/2026-w34), read 16 November 2026.
