📝 摘要
tokenMeter 的 chars/4 估算对 CJK 文本少算约 2.3 倍。
#10153 · SparkElf · 2026-10-18
tokenMeter 的 chars/4 估算对 CJK 文本少算约 2.3 倍。
Summary `tokenMeter`'s fixed `CHARS_PER_TOKEN = 4` heuristic systematically under-prices CJK content. Measured against a live OpenAI-compatible route (`deepseek-v4.1-flash`), the same request over-counts English text by ~1.4x and under-counts Chinese text by ~2.3x. On a real agent session whose surf……
节选自讨论正文开头,完整内容见原文