Gemini 3.8 Flash Is Live: Test Search Citations Before Switching

What Google documents for Gemini 3.8 Flash

The Gemini API changelog marks Gemini 3.8 Flash generally available on September 2. Google’s model page lists text, image, video, audio, and PDF inputs with text output. The documented input limit is 1,048,576 tokens and the maximum output is 65,536 tokens.

Supported capabilities include Google Search grounding, Google Maps grounding, File Search, function calling, code execution, caching, structured outputs, URL context, and computer use in preview. The page lists low, medium, and high thinking levels. It warns that setting minimal causes an error.

Write hypotheses without writing the result

A model release can justify three hypotheses: 3.8 may select different sources, attach citations to different claims, or change latency and output shape. None is a finding until measured. Avoid labels such as “improved grounding” unless Google explicitly documents the improvement or your controlled evidence supports it.

Our earlier Gemini 3.7 Flash grounded-answer test provides the baseline protocol. Reuse its query families and scoring definitions so the model ID is the main changed variable. The Gemini citation-span guide explains why a returned source list is not the same as claim-level support.

Run a matched grounding comparison

Download the Gemini 3.8 grounding comparison ledger (CSV). Every starter row is marked EXAMPLE-REMOVE and contains no observed result.

1. Freeze a query set before looking at model outputs. Include recent facts, multi-claim questions, source conflicts, and stable reference facts.
2. Send the same prompt, grounding configuration, locale, and application instructions to 3.7 and 3.8.
3. Save the exact model ID, UTC time, response, grounding metadata, URLs, latency, and error state.
4. Atomize each answer into checkable claims. Mark a claim supported only when a cited source directly supports it.
5. Report citation validity and claim support separately. A URL can resolve while failing to support the nearby sentence.
6. Repeat enough times to detect output variability; do not cherry-pick the cleanest response.

If the query is time-sensitive, capture the source page at test time. A correct answer can appear wrong after the source changes, and a stale answer can appear right after a later edit.

Score citations at the claim level

Use at least four states: supported, contradicted, unverifiable, and not citation-required. The denominator should be factual claims that require evidence, not paragraphs or sentences. One sentence can contain several claims, and one citation can support only part of it.

For visibility research, add whether your domain was cited, whether it was linked, and whether the linked URL matched the supporting page. Do not treat domain presence as recommendation. Our cited-versus-recommended protocol keeps those states distinct.

Migrate with a reversible gate

Start with shadow traffic or an offline replay. Define thresholds before comparing models: maximum contradiction rate, minimum claim-support rate, acceptable latency, parser success, and cost envelope. A candidate that is faster but loses citation metadata may be unacceptable for a research product; a candidate with stronger evidence but higher latency may suit a batch workflow.

Keep the previous model configuration available during the evaluation window. Log the model ID with every answer. If downstream code assumes a particular grounding shape, include contract tests for missing sources, multiple citations, redirects, and zero-citation answers.

Do not send sensitive or licensed material merely because the context window is large. Apply the same data-minimization, retention, and access rules used for the earlier model.

A defensible release decision

Approve the migration only if the candidate meets the predefined evidence and operational thresholds on representative work. Record failures as well as passes. If results differ by query type, route selectively instead of averaging away a serious weakness.

The strongest conclusion available before testing is narrow: Gemini 3.8 Flash is a stable GA model that supports Search grounding. Everything about comparative citation quality remains a question for the ledger.

Source, method, and update note

Primary sources: Google’s Gemini API changelog and Gemini 3.8 Flash model page, checked September 3, 2026.

Method: SearchEngineAnswer mapped documented capabilities to an unrun matched-model protocol. We report no comparative quality, citation, latency, or cost result.

Recheck trigger: Update if Google changes the stable model ID, capability table, context limits, thinking levels, grounding metadata, or deprecation schedule.