Claude Pro subscribers using non-English prompts are hitting a sharp wall with Opus 4.7: the model appears to consume session tokens (the monthly usage budget Anthropic allocates per plan tier) at a dramatically higher rate than its predecessor when the input language isn't English.
One documented case involves a German-language prompt that Sonnet completed without issue and Opus 4.6 finished in roughly five minutes. Opus 4.7, starting from zero usage, burned through the entire session limit in seconds on the same task. The user confirmed it was a fresh session with no prior usage, ruling out an already-depleted quota.
The most likely explanation is that Opus 4.7 handles non-English input by translating internally to English before processing, then translating back - a pattern common in large language models trained heavily on English text. That extra round-trip generates far more internal computation than processing English directly, and on a model as large as Opus 4.7, the difference appears to be extreme.
Anthropics hasn't publicly acknowledged the issue or confirmed whether this is intended behavior. For now, the practical takeaway is straightforward: if you're working in any language other than English and running Claude through the web interface or API on a usage-capped plan, Sonnet is the safer choice. It handles non-English tasks reliably and without the apparent token penalty. Reserve Opus 4.7 for English-only workloads until Anthropic clarifies what's happening.