Opus 4.7's New Tokenizer: What's the Real Cost

Anthropic added a new tokenizer to Claude Opus 4.7 but kept the input $5/M, output $25/M pricing unchanged. However, the same text is now counted as more native tokens, which can change the actual billed amount. OpenRouter compared over 1 million text-only, non-cancelled requests from a switcher cohort whose primary model changed from 4.6 to 4.7.

The baseline was OpenRouter's own tokenizer, QuadChars, which counts 4 printable ASCII characters as 1 token and treats each non-ASCII character as a separate token, isolating only the difference from the native count. This approach separates out the effect of the tokenizer change alone, not changes in model content.

• < 2K: native token ratio increased by about 45%
• 2K–10K: increased by about 42%
• 10K–25K: increased by about 34%
• 25K–50K and 50K–128K: increased by about 32%
• 128K+: increased by about 33%

For longer prompts, prompt caching absorbed a significant portion of the increased tokens. Above 25K, most of the increase went into the cache, and in the 128K+ range, 93% of the additional tokens were handled by the cache. Conversely, requests under 2K had a low cache hit rate, so the buffering effect was minimal.

Response length also changed. Under 2K, 4.7's median completion was 62% shorter, but at 10K and above, it was 13-30% longer. As a result, the actual average cost rose by 12-27% in the 2K and above range, with per-bracket figures of 2K–10K +27.2%, 10K–25K +25.2%, 25K–50K +21.3%, 50K–128K +11.9%, and 128K+ +15.3%. Under 2K was actually cheaper, at -1.6%.

In other words, the impact of the new tokenizer varied greatly depending on the workload. Short requests became cheaper thanks to shorter responses, while real-world prompts of medium length or longer showed a clear increase in billed amounts even after the cache absorbed part of the increase.