Temperature Control Removed

According to the Claude Opus 4.7 migration documentation, setting temperature, top_p, or top_k to a value other than the default results in a 400 error.

The recommended workaround is to omit these parameters from the request entirely, and the guidance says to steer model behavior via the prompt instead. The author criticizes this change as a major blow to real-world use, since it makes it impossible to control sampling for purposes like code generation.

The post also notes that thinking content is excluded from the response by default; it remains present in streaming, but the thinking field must be explicitly opted into to be visible. This change applies without errors, and is said to slightly reduce latency in exchange.

The core complaints in the post are as follows:

• Low-temperature sampling is useful not so much for determinism as for suppressing low-probability tokens, and that control has now disappeared.
• If the default behaves like temperature = 1, output variability is expected to increase.
• For tasks like long code generation where continuity matters, random sampling could hurt quality, the author argues.