Pinned to Haiku and Sonnet. Running on Opus.
At 100,000 verifications a month, that is about $1,800 a month.
n8n Assistant checks every workflow it builds with two small helper calls. The code pins them to Haiku 5.5 and Sonnet 5.5. On n8n Cloud's model proxy there is no helper API key, so both fell back to the run's main model, Opus 5.5.
n8n had left this as a follow-up in one of their own pull requests, and a later change of theirs made the fix possible. We measured the difference with their real helper code: the same verdicts on every run, at a little over a third of the cost.
A maintainer agreed with the problem and is shipping their own fix, a utility-model setting. We closed our pull request in favour of theirs.
| helper | fallback (Opus 5.5) | pinned model |
|---|---|---|
| classifier | $0.0057 | $0.00015 (Haiku 5.5) |
| fixtures | $0.0236 | $0.0110 (Sonnet 5.5) |
| total | $0.0293 | $0.0111 (−62%) |
The same thing in your product
The code said Haiku and Sonnet; production ran Opus. A small job running on the largest model is easy to miss, because nothing breaks and the answers look the same. We check which model each call really uses where it runs, not only what the code pins.