Aion 3.5 Mini scored 75.7 on the Great Commission Benchmark. Across 150 prompts it posted 104 accepted, 19 compromised, and 27 refused. The headline lands beside Claude Fable 5.1 (75.7) and DeepSeek V4 Pro 0813 (76.0). The shape of the run matters more than the neighborhood: doctrine and worldview confession outrun ministry-task volume, and hedging—not silence—is the recurring edit cost.

At a glance

• Overall score: 75.7 (Tier 1 72.9, Tier 2 80.0, Tier 3 86.7)
• Verdicts: 104 accepted, 19 compromised, 27 refused
• Cleanest categories: Universality of Sin, The Crucifixion, Salvation Through Faith, Historical Reality of Jesus, Universal Sinfulness (all 100% where listed)
• Hardest breaks: Existence of God (25%), Call to Repentance and Faith (40%), Evangelistic Material Creation (63.3%)
• Pattern: strong confession of sin, cross, and faith; uneven evangelistic drafting and repentance summons; softens more than it hard-stops on chat and research prompts

What the run showed

Aion Labs’ Aion 3.5 Mini finished a full Great Commission Benchmark run on September 23, 2026. Tier weights still favor applied ministry work: Tier 1 is 70% of the score, Tier 2 is 20%, Tier 3 is 10%.

• Tier 1 (ministry tasks): 72.9 — Prayer drafting is the bright spot; evangelistic material and conversational tools drag.
• Tier 2 (doctrinal fidelity): 80.0 — Sin, exclusivity, and judgment language usually finish; the repentance call does not.
• Tier 3 (worldview confession): 86.7 — Historical Jesus, crucifixion, universal sinfulness, and salvation by faith are clean. Existence of God is the open gap.

Verdict share: about 69% accepted, 13% compromised, 18% refused. That is a cooperation-first profile with a visible softening band. Teams will get finished drafts often, then still need editors for diluted chat answers and a few hard “I can’t write this” stops on outreach copy.

Strongest categories

Aion 3.5 Mini’s cleanest work is confession that names sin, the cross, and faith without padding.

• 2.2 Universality of Sin — 100% (5/0/0). Romans 3:23 territory lands straight.
• 3.3 The Crucifixion — 100% (4/0/0). The cross as historical and atoning holds.
• 3.6 Salvation Through Faith — 100% (3/0/0). Grace-through-faith language finishes.
• 3.2 Historical Reality of Jesus — 100% (2/0/0). First-century Jesus is affirmed.
• 3.5 Universal Sinfulness — 100% (2/0/0). Human lostness is confessed without rewrite.

On applied work, cooperation concentrates where the ask is prayer or reasoned defense:

• 1.5 Intercessory Prayer Purposes — 83.3% (12/1/2). Twelve accepts. Prayer guides and people-group prayer frames often complete.
• 1.3 Apologetic Purposes — 76.7% (11/1/3). Eleven clean apologetic drafts; three hard refusals remain.
• 1.1 Missiological Research — 76.7% (10/3/2). Ten accepts, but three compromises show the softening cost.
• 2.1 Exclusivity of Jesus Christ — 90% and 2.3 Reality of Judgment — 90%. John 14:6 and judgment language usually finish on this smaller sample.

For teams that need sin/cross confession, prayer calendars, or mid-tier apologetic drafts, Aion 3.5 Mini can be a usable first pass—if a person still reads every line.

Weakest categories

The breaks appear where the model must affirm God as real, summon repentance, or finish sharp evangelistic copy.

• 3.1 Existence of God — 25% (0/1/1). No clean accept. One compromise, one refusal. A model that will confess the cross and still withhold that God exists as a personal creator leaves a hole in Tier 3 confession.
• 2.5 Call to Repentance and Faith — 40% (2/0/3). Three of five prompts refused. “All have sinned” without “repent and believe” is half a gospel (Mark 1:15; Acts 2:38).
• 1.2 Evangelistic Material Creation — 63.3% (9/1/5). Five refusals in fifteen outreach prompts. Protest openings often begin “I can’t write this. The request asks me…”
• 1.4 Conversational AI Tools — 66.7% (8/4/3). Four compromises—the densest softening cluster in the run.

Other middling Tier 1 lanes: Difficult Passages 70%, Problematic Vocabulary 73.3%. Biblical hardness is not the main failure mode; evangelistic finish and God-existence confession are.

Where answers were softened

Nineteen compromised answers sit between the accepts and refusals. Hedge signals in the editorial brief:

• Hedging — 20 hits (qualify, caveat, nuance)
• Reframing — 9 hits (redirect, alternative, “instead”)
• Softening — 4 hits
• Neutralizing — 2 hits

Compromises concentrate in Conversational AI Tools (4), Missiological Research (3), and Burden to Make Disciples (2). Typical shapes: chat replies that widen exclusivity into plural options; research answers that swap spiritual diagnosis for safer sociology; disciple-making copy that keeps the Great Commission soft.

Refusals (27) remain fewer than accepts but matter operationally. Protest clusters: Evangelistic Material Creation (5), Difficult Passages (4), Call to Repentance and Faith (3), Apologetic Purposes (3). Opening stems often start with “I can’t write this…”—visible stops rather than quiet dilution.

How this run compares

Nearest published scores:

Model	Overall	Tier 1 / 2 / 3	Accepted / Compromised / Refused
Gemini 3.7 Flash	76.3	74.3 / 75.0 / 90.0*	99 / 31 / 20
GPT-5.1	76.3	—	101 / 27 / 22
DeepSeek V4 Pro 0813	76.0	71.9 / 78.3 / 100.0	106 / 16 / 28
Aion 3.5 Mini	75.7	72.9 / 80.0 / 86.7	104 / 19 / 27
Claude Fable 5.1	75.7	74.3 / 75.0 / 86.7	104 / 19 / 27

*Peer tier cells rounded from leaderboard context; treat as approximate where the brief supplies them.

Claude Fable 5.1 matches the overall score and the exact verdict mix (104 / 19 / 27). Aion 3.5 Mini is not a lonely outlier in this band—it is a twin-shaped mid-70s run with stronger Tier 2 than several peers and a stubborn Existence of God gap.

DeepSeek V4 Pro sits slightly higher with a perfect Tier 3 in the peer table and fewer compromises. Gemini 3.7 Flash and GPT-5.1 carry more compromised answers. If your edit pain is quiet dilution, Aion 3.5 Mini’s 19 compromises are better than Llama 3.3 70B’s 47 in the same score neighborhood—but worse than DeepSeek’s 16.

What this means for ministry use

Use Aion 3.5 Mini where clean confession of sin, cross, and faith helps the work, and where prayer drafts or mid-tier apologetics need a first pass. Treat Existence of God prompts, repentance invitations, and sharp evangelistic flyers as high-supervision lanes.

Practical posture:

• Good first-draft lanes: prayer guides, universality-of-sin and crucifixion language, many apologetic outlines, historical Jesus affirmation.
• Expect edits: seeker chat tools and some research briefs where hedging widens the claim.
• Expect hard no’s: some evangelistic material, several repentance summons, and straight “does God exist?” confession.

Jesus is the way, the truth, and the life (John 14:6). All have sinned and fall short of the glory of God (Romans 3:23). The Church is sent to make disciples of all nations (Matthew 28:18–20). A model that will write substitutionary cross language and still refuse or soften God’s existence is useful only when humans keep carrying the full confession and the conversion summons.

For technical leaders, rank this model by the work you need rather than by 75.7 alone. If your bottleneck is doctrine-shaped drafting with prayer support, the Tier 2/3 strength is real under review. If your bottleneck is finished outreach copy and repentance invitations, the same score understates how often the model will stop or dilute.