Contents
An analysis of roughly 1,400 Claude Code sessions found an abrupt change around November 3, 2026, alongside reports of worse coding results. This evidence warrants investigating a behavioral change, but does not establish that Claude Opus 5.5’s underlying reasoning ability declined.
Key Takeaways#
- Shane Melaugh reported that initial reasoning became shorter after November 3, while a metric tracking execution and editing errors rose by 47%.
- In one debugging comparison, the “high” setting reportedly failed to solve the problem after 160 exchanges, while “xhigh” succeeded in five.
- A separate analysis of roughly 458,000 responses by Theo Browne found changes in visible reasoning, but no decline in capability in his environment.
- Visible reasoning is a generated summary, and changes in its wording alone cannot establish a decline in reasoning ability.
What Changed on November 3#
Developer Shane Melaugh began investigating after noticing that Claude Opus 5.5 was struggling with tasks it had previously handled. According to the supplied GIGAZINE article, he used Gemini 4 Argon to analyze roughly 1,400 Claude Code sessions recorded between October 9 and November 7, 2026.
The analysis found an abrupt change in the wording of visible reasoning around November 3, or approximately 3 a.m. on November 4 in Japan. Sentences beginning with “I’m” rose from roughly 20–30% to 61.3%. Expressions such as “I should” and “I need to” fell from roughly 50% to 10.9%.
The reported change also appeared in seven sessions already underway. This makes a service-side change worth investigating, but does not identify which component changed or whether that change worsened task performance.
Evidence on Performance#
Melaugh also compared 14,560 exchanges before the change with 14,871 afterward. According to the report, even with reasoning effort set to “high,” the median number of thinking tokens used before the first action fell from 84 to 54, a decrease of roughly 36%.
A metric tracking script execution and program editing errors rose by 47%. These figures are more relevant to coding reliability than changes in writing style, but their interpretation depends on how the metrics were defined and whether the tasks were comparable.
The analysis also reportedly found no thinking-token output in 40.5% of the exchanges examined. This should not be interpreted as evidence that the model performed no internal reasoning. The relationship between recorded output, visible summaries, and internal computation needs to be established.
Comparing high and xhigh#
The clearest example involved attempts to fix an audio playback bug in the same program. Melaugh reported the following results.
| Setting | Number of exchanges | Reported outcome | Reported cost | |---|---:|---|---:| | high | 160 | Still unresolved after roughly three hours | $13 | | xhigh | 5 | Fix completed | $0.85 |
The maximum reported thinking-token count was 1,941 for high and 12,837 for xhigh, a roughly 6.6-fold difference. In this trial, the setting allowing more reasoning also resulted in a lower total cost.
This provides a useful reason to try different reasoning effort settings when debugging stalls. However, it is insufficient to conclude that xhigh is always 32 times as efficient. With only one bug and potentially different execution paths and session contexts, no general performance ratio can be derived.
Why the Findings Remain Inconclusive#
Citing Anthropic’s documentation, the supplied article explains that users see a reasoning summary generated by another model, rather than a direct record of internal processing. A change to the summarization model could explain a major shift in wording without any change in capability.
Developer Theo Browne reported examining roughly 458,000 Claude Opus 5.5 responses. He found that wording changed at different times for different accounts between October 30 and November 4, which he interpreted as a possible account-by-account rollout of a summarization model. In his data, the average number of thinking tokens per response increased from 369 to 413, with no signs of declining capability.
These findings do not invalidate Melaugh’s experience. Rather, they show why the claim remains unresolved. Observations differ across users, and neither analysis provides a controlled, reproducible benchmark demonstrating a general decline. As of the supplied November 10 article, the cause had not been identified.
Next Step#
Choose one reproducible failing task and run paired tests with high and xhigh. Start from the same code snapshot and prompt, and record whether the fixes pass the same tests, elapsed time, and total cost. Choose the setting based on completed results, and treat changes in visible reasoning as a starting point for investigation.
Source: https://gigazine.net/news/20261110-claude-opus-nerf/
Contents
