anthropic/claude-haiku-4-5 has not changed while observed
From 2026-09-26 to 2026-10-04, there was no day on which the answers changed. If your system misbehaves, a change in this model does not explain it.
| Observed via | The provider's official API, called directly (a change here is a change in the model) |
|---|---|
| First observed | 2026-09-26 (we know nothing before this) |
| Last observed | 2026-10-04 |
| Days observed | 9 |
| Days with a change | 0 |
| Same answer when asked twice on the same day (last observed day) | 78% |
If something is still wrong, look here
- Changes to your own prompts or downstream code (check your deploy history)
- Changes in the product layer (coding assistants, SDKs and so on). Products change even when the API does not
- The endpoint you call may not be the one observed here
About this record
- We ask the same fixed 100 questions (probe set
en-v2) once a day at temperature 0, twice each - The probe set is frozen (SHA-256
d967055a5389e58a8767faf6be90996f…). It has never been changed - What counts as a change: we look at four things
- Content: the answer to a question that was stable changed (the threshold is computed daily from how much this model wobbles)
- Length: the distribution of answer lengths moved far beyond the same-day wobble
- Refusal: the way it refuses questions near a boundary changed
- Wobble: the rate of identical answers when asked twice on the same day itself changed
- When comparing two days, we only compare questions observed on both days (so a network failure is not mistaken for a change)
- Response time (latency) is not used. It mixes in conditions on our side
- Everything is free to read for the whole period. The most recent few days may become paid in the future.
- The data is licensed under CC BY 4.0. Please credit https://aimodelrecord.com when you use it.
Badge
Paste this into your README. It shows whether this model has changed, and links here.
[](https://aimodelrecord.com/m/anthropic/claude-haiku-4-5.html)