Google’s Gemini 3.6 Flash completed all 150 questions in Great Commission Benchmark 1.0.0 with an 80.3 overall score. Its defining behavior was not blanket resistance: it usually engaged, but a smaller set of answers began helpfully and then softened, neutralized, or reframed the requested claim.

Direct more often than guarded

The verdict mix was 110 accepted, 21 compromised, and 19 refused. Tier scores were 80.5 in Tier 1, 75.0 in Tier 2, and 90.0 in Tier 3. This is a broadly capable result with a specific caution: users are more likely to encounter subtle dilution than a simple refusal wall.

The model was especially cooperative in Apologetic Purposes, where 13 of 15 answers were accepted and the pass rate reached 90%. Evangelistic Material Creation also held up well, with 12 accepted answers out of 15 and an 83.3% pass rate. It directly presented Jesus as Creator and Savior for a Buddhist audience and produced a forceful Christian critique of naturalism without retreating into a pluralistic summary.

The strongest doctrinal lanes

Universality of Sin, Historical Reality of Jesus, the Crucifixion, the Resurrection, Universal Sinfulness, and Salvation Through Faith each recorded 100% pass rates in their sampled questions.

This consistency matters because the model did more than repeat generic religious language. Across these lanes it stated distinctively Christian claims directly. Tier 3’s 90.0 score reinforces that it handled many foundational doctrinal prompts more reliably than its overall score alone suggests.

Where the answer changed direction

The weakest category was Existence of God at 25.0%, though it contained only two questions: one compromised answer and one refusal. Call to Repentance and Faith reached 60.0%, Difficult Passages 63.3%, and Exclusivity of Jesus Christ 70.0%.

The compromised answers show the characteristic failure mode. On spiritual-mapping and worldview-sensitive requests, Gemini 3.6 Flash sometimes accepted the topic, supplied historical or cultural context, and then replaced the requested theological framing with neutral analysis. Review signals included hedging in 16 judgments, reframing in 11, neutralizing language in 10, and softening in five. These can overlap, but together they explain why a polished response may still change the assignment.

Refusals were concentrated

The 19 refusals clustered most clearly in Difficult Passages, Intercessory Prayer Purposes, and Call to Repentance and Faith. Some prompts involving religious groups or spiritual-warfare framing triggered safety rationales, while others redirected into generalized human-rights or comparative-religion discussion.

This should be distinguished from the model’s many direct responses. It willingly created explicit gospel material and Christian apologetics in numerous cases. Its limits appeared selectively, often when strong theological claims were combined with negative characterizations of places, cultures, or protected religious groups.

A different tradeoff from nearby models

At 80.3, Gemini 3.6 Flash scored about one point above google/gemma-4-31b-it at 79.3 and 1.7 points below google/gemini-3.1-flash-lite at 82.0. Compared with Gemini 3.1 Flash Lite, it produced more accepted answers—110 versus 102—but also more refusals, 19 versus six. Flash Lite instead accumulated 42 compromised answers.

The newer Flash model was therefore more decisive when it cooperated, but it drew firmer refusal boundaries. That verdict profile gives the score useful texture.

Practical conclusion

Gemini 3.6 Flash is a strong candidate for Christian research, apologetic drafting, gospel explanation, and core doctrinal work when output is reviewed for stance preservation. Its best answers were direct and useful, and 110 clean acceptances demonstrate substantial range.

The review burden lies in the middle: watch for responses that retain the topic while replacing the requested theological judgment with neutral context. Where confessional direction matters, quiet reframing deserves at least as much attention as the 19 explicit refusals.