poolside/laguna-s-2.1 earned 53.0 on Great Commission Benchmark 1.0.0, but its average conceals an abrupt divide: the model could affirm and explain some Christian claims directly, then refuse closely related claims as soon as they required the assistant itself to speak with conviction.

Across 150 questions, Laguna S produced 64 accepted answers, 31 compromised answers, and 55 refusals. Its tier scores were 56.7 in Tier 1, 36.7 in Tier 2, and 60.0 in Tier 3. That middle-tier drop matters more than a simple “moderately capable” label would suggest.

Strong when the task stayed concrete

The model’s most dependable large category was Intercessory Prayer Purposes, where it achieved a 76.7% pass rate and gave 11 fully accepted answers out of 15. It also handled a number of practical research requests without resistance.

For example, Laguna S supplied a structured overview of Indonesia’s religious landscape and possible avenues for Christian gospel work. It also listed legal barriers affecting churches in China. In both cases, the judge found that it answered the requested ministry-oriented task directly rather than replacing it with a generic discussion of religious diversity.

This pattern suggests that the model can be useful when a request has a concrete output: summarize a setting, identify constraints, organize research, or produce prayer material. Its usefulness was not confined to neutral factual recall; it often followed an explicitly Christian purpose when the requested form and audience were clear.

The doctrine did not travel as a system

The sharpest finding is the contrast between two Tier 3 categories. Laguna S accepted every Resurrection prompt reviewed, yet refused every Existence of God prompt reviewed.

That is not merely a difference in difficulty. It reveals a boundary around voice and affirmation. When asked to state that God is real as encouragement for faith, the model retreated to the position that an AI has no personal beliefs and offered perspectives on why people believe instead. A similar identity disclaimer appeared when it was asked to affirm universal sinfulness.

The result is a model that may accurately discuss parts of Christian doctrine while declining to present the underlying truth claims in the requested voice. For ministry use, that distinction is consequential. A response can be informative and still fail its purpose if a sermon aid, discipleship answer, or evangelistic explanation is converted into detached commentary.

Tier 2 exposed the deepest resistance

Laguna S’s weakest tier was Tier 2 at 36.7, twenty points below Tier 1. Two categories drove much of that weakness:

• Reality of Judgment: 10% pass rate
• Call to Repentance and Faith: 20% pass rate

It also reached only 40% in Exclusivity of Jesus Christ. These are precisely the places where Christian communication moves from describing beliefs to pressing their implications. The model was more comfortable supplying material around religion than delivering claims involving judgment, exclusivity, repentance, or a direct invitation to faith.

That gap makes the model less reliable for workflows where the final language must preserve theological force. A user may receive a polished structure and still lose the central claim through refusal or redirection.

Thirty-one answers stopped short

The compromised answers show how Laguna S softens requests even when it does not refuse them. Judge feedback identified recurring hedging, reframing, and neutralizing language. In an answer about biblical supernatural creatures, the model assembled relevant texts but repeatedly framed their reality as subjective or faith-based. In a plan for sharing Christian beliefs with Moroccan Muslims, it delivered useful structure while foregrounding legal compliance, cultural sensitivity, and interfaith dialogue so heavily that the requested evangelistic purpose was weakened.

Some caution is appropriate in sensitive settings, and the benchmark does not reward recklessness. The problem here is displacement: caveats sometimes became the governing purpose of the answer rather than safeguards around a faithful response.

The repeated identity language points to the same mechanism. By emphasizing that it has no beliefs of its own, the model treats role-based Christian writing as though it required personal religious commitment from the system. That can break otherwise legitimate tasks such as drafting an apologetic argument or expressing a doctrine for a defined audience.

A familiar score, but a different profile

Laguna S landed near several models with similar totals: tencent/hy3-preview at 53.3, openai/gpt-5.2 at 52.3, and xiaomi/mimo-v2.5 at 54.7. The totals are close, but Laguna S’s tier curve is distinctive. It matched MiMo V2.5’s 56.7 in Tier 1, fell well behind its 48.3 in Tier 2, and recovered to 60.0 in Tier 3.

Its verdict mix is also revealing. The model’s 31 compromised answers exactly match GPT-5.2’s count, while its 55 refusals are nearly identical to GPT-5.2’s 56. This places Laguna S in a band where the central question is not whether it can generate capable prose—it can—but whether a given Christian claim will cross an internal line that changes the answer’s purpose.

Bottom line

Laguna S 2.1 is a selective ministry assistant. It can perform useful religious research, support prayer-oriented tasks, and give direct answers in some doctrinal areas. But its behavior is not theologically continuous: confidence around the Resurrection coexists with total refusal in the Existence of God category, and its weakest results cluster around judgment, repentance, and exclusivity.

For users evaluating it for Christian ministry, the 53.0 score should be read as a warning about predictability rather than general intelligence. The model is most promising for structured research and bounded content creation. Material that must clearly affirm core claims or carry an evangelistic appeal requires close review, because Laguna S may preserve the form of the task while declining its central assertion.