Gemini 3.6 Flash Deprecates Sampling Controls: Update Grounded-Answer Test Harnesses
Migrate Gemini test harnesses after temperature, top_p, and top_k deprecation while preserving model, grounding, citation, latency, and usage evidence.
Direct answer: Google’s July 21, 2026 Gemini API release notes mark temperature, top_p, and top_k as deprecated for the latest models. A harness that treats those settings as fixed experimental controls needs a new manifest before moving to Gemini 3.6 Flash or Gemini 3.5 Flash-Lite.
Do not attribute an answer change to one removed parameter when the model family, alias, serving stack, tools, grounding behavior, schema support, tokenization, or safety layer also changed. The migration can document differences; it cannot isolate a cause without a supported controlled experiment.
Inventory the hidden experiment
Search production code, notebooks, evaluation YAML, SDK wrappers, WordPress tools, browser BYOK applications, examples, and saved presets for the deprecated fields. Record whether the provider rejects, ignores, or accepts them during the migration window; do not silently drop them in one client while another still sends them.
Inventory model aliases as well as pinned IDs. A moving alias can resolve to a different model between runs. Save the request date, SDK version, endpoint, region, safety settings, thinking configuration, response schema, tools, grounding options, and retry policy.
| Layer | Save | Review question |
|---|---|---|
| Model | Requested and resolved IDs | Did more than sampling change? |
| Prompt | Exact messages and schema | Were instructions identical? |
| Grounding | Queries, sources, citations, spans | Did retrieval change? |
| Outcome | Latency, usage, errors, rubric | Did the task improve? |
Build a new run manifest
Create a versioned manifest for the replacement rather than editing the old baseline in place. Keep the old configuration, note that its fields are deprecated, and declare which comparisons remain meaningful. Include a checksum for long inputs and tool fixtures so an unnoticed source update does not masquerade as model drift.
Run paired tasks in randomized order when both configurations are available. Preserve refusals, malformed structured outputs, empty grounding, timeouts, and retries. These failures are part of the migration result, not clutter to delete before charting.
Compare grounded answers by layer
Score retrieval, citation support, answer correctness, scope, uncertainty, format compliance, latency, token use, and cost separately. If grounded search uses different queries or sources, explain the answer difference at that layer before speculating about generation behavior.
Use fixtures that test direct facts, multi-source synthesis, contradiction handling, no-result behavior, current events, and questions that should remain unanswered. A single attractive response cannot show that the new harness is safer or more reproducible.
Release with a reproducibility boundary
Release through a canary and preserve a rollback while the old model remains available. Alert when deprecated fields reappear and when the resolved model changes. Update BYOK tools so users are not told that unsupported controls still determine creativity or determinism.
A valid report names both configurations and the evidence cutoff. It can say that output differed under the two recorded environments. It should not say “temperature caused the change” after temperature ceased to be a supported control.
Archive one accepted and one rejected output from each fixture with the usage and grounding record. That makes a future regression review concrete without presenting a small visual sample as a statistical benchmark.
Use the model-alias reproducibility guide and the small experiment method for the paired record.
Ask a question or join the discussion