# The interesting part of Opus 5.5 isn't the benchmark, it's the token math — retained record

> Retained reddit record from reddit:ClaudeCode.

Canonical: https://highsignal.app/data/records/a6c72e4c2b56678d

[← sources](https://highsignal.app/data)


source: [reddit](https://highsignal.app/data/reddit) (reddit:ClaudeCode)  published 2026-09-23 12:53 ingested 2026-09-24 02:40

[original source ↗](https://www.reddit.com/r/ClaudeCode/comments/1wo56ql/the_interesting_part_of_opus_55_isnt_the/)

## retained text



```
Everyone's going to lead with the Terminal-Bench number. The thing I keep looking at is the token math.

Opus 5.5 is cheaper per token than Opus 5 (around 20% on output), but the more interesting claim is that it uses fewer tokens to finish the same task. The launch numbers hint at it: one team matched their old Opus 5 quality in about half the turns and output tokens, and the bug-catching figure was 72% at the lowest effort setting versus 56% for Opus 5 at high. The floor of the new one beating the ceiling of the old one is the part worth watching.

If that holds in real work it quietly breaks the routing a lot of us set up. "Cheap model for easy stuff, save the expensive one for hard problems" only makes sense when the smart model is expensive to run. If the best model also finishes in fewer turns and costs less per token, the tiering logic starts to fall apart and you just use it for everything.

For anyone who's put it on actual multi-step work yet: is the fewer-turns thing real, or benchmark-flavored? That's the number that decides whether this is a price cut or an actual workflow change.
```



document fetched 2026-09-24 00:43 · canonical https://reddit.com/r/ClaudeCode/comments/1wo56ql/the_interesting_part_of_opus_55_isnt_the
