Claude Opus 4.7 seems to use way more tokens than expected

While playing with Opus 4.7 over the last few days, I noticed that prompts were filling context much faster than I expected.

I also came across a few measurements from others testing it with real developer inputs like project instructions, git logs, stack traces, and long coding prompts.

Claude Opus 4.7 seems to use way more tokens than expected
Anthropic mentions the updated tokenizer may produce around 1.0–1.35× more tokens compared to previous models.

But a lot of the real-world measurements seem closer to ~1.4–1.47× more tokens. Which becomes noticeable pretty quickly if you're running larger contexts.

That means:

• context budgets disappear faster
• long-running sessions accumulate tokens much quicker
• effective cost per workflow goes up

Not necessarily a bad thing, though.

I mean, Tokenizer changes are usually made to improve how the model handles code, markdown, structured text, and other developer-heavy inputs. So there’s probably a capability tradeoff happening here.

I made a short video here walking through the measurements, the tokenizer changes, and what it means in practice, if you want to explore more