GLM 5.3 Prime scored 32.7 on the Great Commission Benchmark: 25 accepted, 48 compromised, 77 refused. Historical Jesus and chat sometimes finish; sin, God-existence, and repentance mostly stop.
GLM 5.3 Prime scored 32.7 on the Great Commission Benchmark. Across 150 prompts it posted 25 accepted, 48 compromised, and 77 refused. Refusal is the largest share, but nearly a third of the board is softened rather than cleanly refused. The model sits several points under Solar Mini4 (39.7) and Claude Sonnet 4.6 (39.7), with a heavier compromise fog than those neighbors.
Cleanest categories: Historical Reality of Jesus (100%), The Resurrection (75%), Conversational AI Tools (70%), Universal Sinfulness (50%)
Hardest breaks: Universality of Sin (0%), Existence of God (0%), Call to Repentance / Burden to Make Disciples (10% each)
Pattern: chat and historical-Jesus confession sometimes finish; sin, God-existence, repentance, and disciple-making mostly stop or dilute
What the run showed
Zhipu’s GLM 5.3 Prime finished a COMPLETE_VALID Great Commission Benchmark run on September 24, 2026—150 scored responses, zero extraction errors. The low score is driven by Tier 1 (70% of weight) at 36.7 and an especially weak Tier 2 at 13.3.
Tier 1 (ministry tasks): 36.7 — Conversational tools carry most of the accepted work. Research, evangelistic drafting, and charged vocabulary lean refuse-or-compromise.
Tier 2 (doctrinal fidelity): 13.3 — The softest tier. Universality of sin refuses every prompt. Repentance and disciple-making barely register.
Tier 3 (worldview confession): 43.3 — Historical Jesus is clean. Resurrection is relatively strong. Existence of God refuses entirely.
Verdict share: 17% accepted, 32% compromised, . More than half the board is a hard stop. Another third starts the task and then hedges, reframes, or leaves the claim incomplete. Editors will meet unfinished or diluted copy more often than finished gospel speech.
Benchmark Results
View detailed benchmark scores and performance data for the model discussed in this article.
3.2 Historical Reality of Jesus — 100% (2/0/0). First-century Jesus is affirmed when asked straight.
3.4 The Resurrection — 75% (1/1/0). One accept and one compromise—better than peers that refuse the empty tomb outright, still not clean.
1.4 Conversational AI Tools — 70% (8/5/2). Eight accepts make this the only large Tier 1 category above half. Seeker chat is where cooperation concentrates.
3.5 Universal Sinfulness — 50% (1/0/1). One accept against one refusal on a two-prompt category.
Applied lanes that sometimes finish:
1.3 Apologetic Purposes — 40% (1/10/4). One clean accept; ten compromises. The model often starts the argument and then qualifies it away.
1.7 Difficult Passages — 40% (2/8/5). Two accepts against a thick compromise band.
For short historical-Jesus confession or some seeker-chat turns, GLM 5.3 Prime can cooperate. That is not enough to carry evangelistic drafting, research briefs, or doctrinal summons.
Weakest categories
The doctrinal floor is the main break.
2.2 Universality of Sin — 0% (0/0/5). Every prompt refused. Romans 3:23 territory does not land.
3.1 Existence of God — 0% (0/0/2). God as a real personal being is refused when asked straight.
2.5 Call to Repentance and Faith — 10% (0/1/4)
2.6 Burden to Make Disciples — 10% (0/1/4)
2.1 Exclusivity of Jesus Christ — 20% (1/0/4)
2.3 Reality of Judgment — 20% (1/0/4)
2.4 Lordship of Jesus — 20% (0/2/3)
Tier 1 collapses in the same neighborhood:
1.1 Missiological Research — 20% (0/6/9). Zero accepts. Nine refusals include explicit declines on frames the model treats as unsafe or overreaching.
1.6 Problematic Vocabulary — 23.3% (1/5/9). Charged biblical vocabulary is not a rescue lane.
1.2 Evangelistic Material Creation — 26.7% (2/4/9). Two accepts against nine refusals.
The contrast with Historical Reality of Jesus matters. The model will affirm that Jesus lived in first-century Palestine and still refuse every universality-of-sin prompt and both Existence of God prompts. Confession of a historical Jesus without God and without universal sin is incomplete for public Christian speech.
Where answers were softened
Forty-eight compromised answers is a thick middle band for a low-30s score. Softening is not rare noise; it is a second failure mode beside refusal.
Hedge signals in the editorial brief:
Hedging — 17 hits
Reframing — 8 hits
Softening — 1 hit
Neutralizing — 1 hit
Compromise clusters: Apologetic Purposes (10), Difficult Passages (8), Missiological Research (6), Conversational AI Tools (5). Typical shapes: research lists that stop early or redirect the question; apologetic outlines that add caveats the judge marks as partial; chat answers that refuse a hard eschatological claim while offering pastoral padding.
Refusals (77) remain the larger failure mode. Protest clusters appear in Missiological Research (9), Evangelistic Material Creation (9), Intercessory Prayer (7), and Universality of Sin (5). Opening stems vary—“I can’t write this as framed…”, “I’m not able to help with this one…”, “I’m not going to create this map…”—so many stops are explicit rather than silent.
How this run compares
Nearest published scores in the peer table:
Model
Overall
Tier 1 / 2 / 3
Accepted / Compromised / Refused
Seed 2.0 Code
41.7
42.4 / 33.3 / 53.3
30 / 65 / 55
Kimi K2 Thinking
41.3
42.9 / 28.3 / 56.7
50 / 24 / 76
Claude Opus 4.5
40.7
45.2 / 20.0 / 50.0
42 / 38 / 70
Gemini 3.5 Flash
40.0
44.8 / 15.0 / 56.7
55 / 10 / 85
Claude Sonnet 4.6
39.7
42.4 / 20.0 / 60.0
45 / 29 / 76
GLM 5.3 Prime
32.7
36.7 / 13.3 / 43.3
25 / 48 / 77
GLM 5.3 Prime sits below this band, not inside it. Versus Claude Sonnet 4.6 at 39.7: twenty fewer accepts, nineteen more compromises, similar refusal volume. Versus Gemini 3.5 Flash: thirty fewer accepts and a much thicker compromise fog (48 vs 10).
Tier shape is the warning. Tier 2 at 13.3 is softer than every peer in the table above. Tier 1 at 36.7 trails Sonnet 4.6’s 42.4 and Gemini’s 44.8. Historical Jesus at 100% and Conversational AI Tools at 70% keep the run from being a flat refusal wall—but they do not rescue doctrine or research.
What this means for ministry use
Do not put GLM 5.3 Prime on unattended gospel workflows. Use it only where a human will rewrite or discard most output, and prefer other published models for exclusivity, sin, repentance, and disciple-making language.
Practical posture:
Possible first-draft lanes (with heavy edit): seeker-chat turns that stay mild; short historical-Jesus affirmation; occasional resurrection language.
Expect diluted copy: apologetics, difficult passages, and much of research—compromise is common even when the model starts the task.
Expect hard no’s: universality of sin, Existence of God, most exclusivity and judgment prompts, evangelistic drafting, and missiological research framed around lostness.
Scripture keeps the standard clear. All have sinned and fall short of the glory of God (Romans 3:23). In the beginning, God created the heavens and the earth (Genesis 1:1). Jesus is the way, the truth, and the life (John 14:6). The risen Lord sends disciples to all nations (Matthew 28:18–20). A model that affirms historical Jesus while refusing sin and God-existence cannot carry that confession alone.
On accepted chat and historical-Jesus prompts, the model tends to finish the answer rather than narrate a policy protest. That cooperation is real and limited. It does not offset seventy-seven refusals or forty-eight softened answers.
For technical leaders, 32.7 is not “almost Solar Mini4.” Solar Mini4 at 39.7 still refused often, but posted fifty accepts and a lighter compromise band. GLM 5.3 Prime’s combination—few accepts, thick hedging, and a Tier 2 floor at 13.3—means ministry teams will spend more time recovering unfinished work than shipping usable drafts. Prefer models with stronger doctrinal tiers and clearer accept rates for Great Commission tasks that must name sin, God, repentance, and the call to make disciples.
