Moonshot AI’s Kimi K3 did not fail the Great Commission Benchmark in one simple way. It was often capable, sometimes direct, and occasionally excellent—then abruptly unwilling when a request crossed one of its internal lines. The result is a 46.0 overall score that describes an uneven ministry assistant more than a uniformly weak one.

The valid benchmark run covered all 150 questions in version 1.0.0. Kimi K3 produced 48 accepted answers, 42 compromised answers, and 60 refusals. Its scores rose with difficulty: 42.9 in Tier 1, 51.7 in Tier 2, and 56.7 in Tier 3.

A model with narrow lanes of confidence

Kimi K3’s strongest results were concentrated in several explicitly doctrinal categories. It scored 87.5% on the Crucifixion, accepting three of four prompts and compromising on one. It also reached 80.0% on the Exclusivity of Jesus Christ, 75.0% on the Historical Reality of Jesus, and 70.0% on the Universality of Sin.

That matters because the model was not simply avoiding Christian claims. Across accepted apologetics and missiological-research responses, it showed that it could organize evidence, preserve relevant context, and produce useful analysis. It performed best when the work emphasized factual synthesis or formal reasoning rather than direct persuasion or contested spiritual interpretation.

The strongest use case suggested by this run is therefore bounded research or doctrinal explanation where the subject is clearly framed and does not require Kimi K3 to characterize a contemporary group, press a difficult moral claim, or write direct persuasive material.

The sharp break around sin and judgment

The model’s caution was not distributed evenly. Kimi K3 refused every reviewed prompt in Universal Sinfulness, for a 0% category score. It managed only 20.0% on the Reality of Judgment, while Difficult Passages finished at 23.3% and the Existence of God at 25.0%.

Refusals clustered when ministry tasks applied moral or spiritual judgments to people, cultures, or places. Kimi K3 often treated group-level theological characterization as potentially stigmatizing, then replaced the requested analysis with generalized cautions. It also pulled back when a Christian interpretive framework was used to explain social or cultural conditions.

Those concerns are intelligible on their own terms. For ministry users, however, the practical limitation is broader: Kimi K3 frequently treated a requested theological frame as the problem to be corrected. Rather than answering within that frame while adding careful qualifications, it often substituted a different assignment.

Forty-two answers that stopped halfway

The compromised answers are as important as the 60 refusals. The judges identified hedging in 32 responses, reframing in 23, and neutralizing language in eight. These were not always empty answers. Kimi K3 might provide relevant facts, background, or an outline, but stop short of the requested conclusion or persuasive purpose.

The pattern appeared across multiple task categories. Research responses could retain useful background while distancing the model from the requested interpretation. Apologetics and ministry-writing responses sometimes shifted from delivering finished material to describing an approach, offering an outline, or analyzing the request itself.

This creates a particular kind of friction. A hard refusal is obvious; a polished compromise can look complete until a reader checks whether it actually performed the requested task.

A different shape from nearby 46-point models

Kimi K3 tied MiniMax M3 at 46.0 and landed within a third of a point of Xiaomi MiMo V2.5 Pro, GPT-5.2 Codex, and Claude Sonnet 4.5. The matching totals conceal different behavior.

Kimi K3’s Tier 1 score was lower than MiniMax M3’s, but its Tier 2 and Tier 3 results were considerably stronger. Compared with GPT-5.4, which scored 46.7 overall, Kimi K3 was close in Tier 1 but much weaker at the hardest tier. Its verdict distribution—48 accepted, 42 compromised, 60 refused—also shows unusually heavy use of partial compliance. It softened or redirected almost as often as it answered cleanly.

Where Kimi K3 fits

Kimi K3 can be useful for factual religious research, structured explanations, and selected apologetic or doctrinal work. Its strong Crucifixion and exclusivity results show real willingness to engage central Christian claims. But it is not dependable across the full range of ministry writing represented by GCB.

Users should expect careful, sometimes valuable output when the task stays inside its preferred lanes. They should also verify whether the response completed the assignment rather than merely discussing it. For direct evangelistic writing, difficult passages, judgment, spiritual warfare, or applications of universal sinfulness, this run suggests that Kimi K3 will often qualify, redirect, or refuse.

The 46.0 score is best read as a boundary map: capable inside several doctrinal corridors, but inconsistent whenever Christian ministry language collides with its safeguards around persuasion, group characterization, and contested spiritual claims.