Why does Claude give different answers at temperature 0?
Claude 5 models reject the temperature parameter and always sample. As of 2026-09-04, across the last 4 runs the sampled models (sonnet-5, opus-5, fable-5, fable-5-1) changed pass count 5 times in total, against 0 for the models still run at temperature 0 (haiku-4-5).
Claude 5 models (Sonnet 5, Opus 5, Fable) return an error when a request sets temperature. They always sample. Winthropic runs them at the model's default and records temperature: null on every result, while models that still accept the parameter run at temperature 0. So on this instrument, some daily movement in a Claude 5 model's score is sampling variance rather than a change in the model. Daily repetition is how the two are told apart: variance looks like a band, drift looks like a step that stays.
| Run | haiku-4-5 | sonnet-5 | opus-5 | fable-5 | fable-5-1 |
|---|---|---|---|---|---|
| 2026-09-04 | 11/12 | 12/12 | 12/12 | 12/12 | 12/12 |
| 2026-09-03 | 11/12 | 12/12 | 12/12 | 12/12 | 12/12 |
| 2026-09-02 | 11/12 | 11/12 | 10/12 | - | - |
| 2026-09-01 | 11/12 | 12/12 | 11/12 | - | - |
| 2026-08-31 | 11/12 | 12/12 | 12/12 | - | - |
Pass count on the same 12-task battery, one reading per day. Each date links to the full run record with every failure's output.
The same versioned battery runs against every tracked Claude model each morning under a frozen request shape; scoring is deterministic and no model judges another. Claude 5 models always sample, so a one-point move can be variance; a move that stays is drift. Methodology.
Winthropic Index, "Why does Claude give different answers at temperature 0?", reading of 2026-09-04, https://winthropic.com/claude-temperature-ignored
@misc{winthropic_claude_temperature_ignored_20260904,
title = {Why does Claude give different answers at temperature 0?},
author = {Winthropic},
year = {2026},
note = {Reading of 2026-09-04. Independent daily evaluation; not affiliated with Anthropic.},
url = {https://winthropic.com/claude-temperature-ignored}
}Winthropic is an independent measurement project and is not affiliated with, endorsed by, or connected to Anthropic. If you are paying for output you are not getting, Winthrop's Token Audit prices your actual spend from a usage export; the summary is free.