LLMs, GenAI, RAG, agents, evaluation · insight-1516
Check which sampling parameters each model accepts before running a temperature study
- Evidence type
- documented fact
- Difficulty
- easy
- Roles
- Applied AI, Data scientist, ML engineer
- Industries
- cross-industry
Summary: situation, action, reason, boundary
Situation
An evaluation harness sweeps temperature from 0 to 1 across hosted models, including one released this quarter.
Action
Read the current parameter reference per model, probe each endpoint with every setting and assert on the response, record the effective settings in the run log, and use prompt or format constraints for variability control where sampling controls are unavailable.
Reason
Anthropic's reference marks temperature, top_p and top_k as deprecated, says models released after Claude Opus 4.6 reject values other than a compatibility default, and notes that temperature 0.0 is not fully deterministic.
Boundary
Cross-model evaluations and any system that stores decoding settings with prompt versions.
Symptom or risk
Some requests are rejected, other hosts ignore the setting, so the sweep compares identical configurations or models that are not treated alike.
Exceptions
Self-hosted models, which accept the full set of sampling controls.
Trade-offs
Fewer knobs mean variability must be handled through repeated runs and output constraints.
Pitfalls
Reporting a temperature study whose settings were never confirmed as applied.
Lessons and concepts
Claims and their sources
Anthropic's Messages API reference marks temperature, top_p and top_k as deprecated, states that models released after Claude Opus 4.6 reject non-default values with a 400, and notes that results are not fully deterministic even at temperature 0.0 (as read on 2026-10-12).