What Broke in Claude Code: An Analysis of Three Causes

For several weeks, developers complained that Claude Code had started working worse: the model forgot context, repeated itself, and answered more briefly than usual. Anthropic initially found no confirmation, but later published an analysis and acknowledged three separate causes. The story is useful not so much for its details as for its conclusion: a product can change without a single line in the release notes.

Three causes that came together

The first was a lower reasoning level. In March, the default setting for Claude Code was lowered from high to medium because some users complained about answers taking too long. The model began thinking less deeply about a task if the developer did not change the setting manually. The company later acknowledged that this slightly reduced its intellectual capabilities, but also decreased latency and made users hit limits less often. The change was fully rolled back on 7 April.

The second was a caching error. The optimization released on 26 March was supposed to clear the reasoning history once after an hour of inactivity. Instead, it erased it on every subsequent turn. In long conversations, the model lost its own previous reasoning, began forgetting details, and went around in circles. It was fixed in version v2.1.116.

The third was a brevity instruction. On 16 April, a requirement was added to the system prompt to keep the text between tool calls within 25 words and the final answer within 100. Measurements showed that programming results fell by approximately three percent. The instruction was also removed.

None of the three changes was presented as a deterioration. Each addressed its own issue: latency, resource usage, or answer concision. Coinciding in time, they produced an effect that users described as “the model got dumber.”

Why “secret degradation” is the wrong term

The company denies intentionally weakening the models and says that the weights were not changed. Judging by the published analysis, that is indeed the case: the surrounding systems and system instructions broke, not the model itself. There is no basis for claiming that a proven secret reduction in quality occurred.

For the user, however, the difference is small. They pay for a result, and the result worsened because of changes they knew nothing about and could not influence. The acknowledgment came after several weeks of complaints.

The Same Model Choice—Different Answers

There is a mechanism that reinforces the same impression. Selecting Opus 5 in the interface does not guarantee that this exact model will process the request. Some requests related to cybersecurity are automatically routed internally to Opus 4.8 within Claude, Claude Code, or Claude Cowork. For requests in the fields of security, biology, and distillation, the model decision is made by an internal classifier.

Anthropic has described this system openly, so it cannot be called hidden. The practical consequence is still unpleasant: two people with the same request on the same plan may receive different models, and the interface will not tell them.

Text watermarks are also being discussed separately—a mechanism that slightly shifts word choice during generation so that generated text can be identified. The idea has drawn criticism: once again, the user does not receive exactly what they ordered.

What This Means for Your Work

Record the tool version if the result is critical to you. The client version number and the date of the run are the minimum data without which it will later be impossible to tell whether the product or your project changed.

Keep a short set of your own checks. Five to ten typical tasks that you run once a week while recording the results will catch degradation before it turns into a vague impression. An impression is not actionable evidence, but two measurements three weeks apart are an argument.

Do not rely on default settings for critical tasks. The reasoning level, answer length, and model choice are parameters that a provider may change for load-management reasons. An explicitly configured setting survives such changes; a default setting does not.