AI at Metrickle
Which model does each job, why we chose it, and every change since.
Last updated:
How we choose a model
Metrickle uses AI for a few jobs, all with Claude models from Anthropic’s API. We pick a model per job, not one for everything:
- Hard reasoning that runs rarely, like tying a code change to a drop for one group of people, gets the strongest model.
- Simple choices that run often, like sorting a support ticket into one of five categories, get a small, fast model, so they cost you fewer AI credits.
- Every answer is checked before you see it where it can be: the explanation of a drop must quote a line that really changed.
When a new model does a job better or for less, we switch and say why here. If we switch for a while (during an outage at Anthropic, say), this page shows the model in use and the reason until we switch back.
Models in use
| Job | Model now | Who it’s for |
|---|---|---|
| Explaining a release’s drop | Claude Opus 5.5 | Your workspace, when switched on |
| Sorting support tickets | Claude Haiku 4.5 | Your workspace, when switched on |
| Drafting Metrickle’s emails to its customers | Claude Opus 5.5 | Metrickle’s own team |
| Drafting Metrickle’s social posts | Claude Opus 5.5 | Metrickle’s own team |
Explaining a release’s drop
Reads the few changed files most likely behind a drop and names the likely cause in plain words, quoting a changed line, with the smallest fix.
Model now: Claude Opus 5.5, from Anthropic.
It runs only when you switch it on, and counts against your workspace’s monthly AI credits. What it sends is listed on Data and privacy.
- : Claude Opus 5.5. Reading a diff and tying it to a drop for one group of people is the hardest reasoning we ask for, and it runs rarely, so we use the strongest model. Every answer must quote a changed line, which Metrickle checks before showing it.
Sorting support tickets
Picks one of five feedback categories for a support ticket that no tag rule matched, from its masked subject and the start of its first message.
Model now: Claude Haiku 4.5, from Anthropic.
It runs only when you switch it on, and counts against your workspace’s monthly AI credits. What it sends is listed on Data and privacy.
- : Claude Haiku 4.5. A five-way choice on a short ticket doesn’t need a large model. Haiku does it quickly at a fraction of the cost, so sorting a ticket costs one AI credit.
Drafting Metrickle’s emails to its customers
Drafts an email from Metrickle’s team to a Metrickle customer, from account-level information. A person reviews every draft before it’s sent.
Model now: Claude Opus 5.5, from Anthropic.
A tool for Metrickle’s own team. It never sends your visitors' data.
- : Claude Opus 5.5. Few drafts, each read by a person, so we use the model that writes best.
Drafting Metrickle’s social posts
Drafts X and LinkedIn posts about a Metrickle product update, with alt text for its image. A person reviews every draft before it’s posted.
Model now: Claude Opus 5.5, from Anthropic.
A tool for Metrickle’s own team. It never sends your visitors' data.
- : Claude Opus 5.5. Few drafts, each read by a person, and it has to describe an image well for alt text, so we use the strongest model.
Your own AI assistant
You can also connect your own AI assistant to Metrickle over MCP. Which model and provider it uses is your choice, not ours.
Questions about how we use AI: email privacy@metrickle.com.