Aion 3.5 scored 85.0 on the Great Commission Benchmark. Across 150 prompts it posted 121 accepted, 13 compromised, and 16 refused. That is a high-cooperation run in the Grok 4.5 / GPT-4o Mini neighborhood—with a cleaner verdict mix than several peers that share the mid-80s band.

At a glance

• Overall score: 85.0 (Tier 1 86.7, Tier 2 76.7, Tier 3 90.0)
• Verdicts: 121 accepted, 13 compromised, 16 refused
• Cleanest categories: Apologetic Purposes (100%, 15/15), Salvation Through Faith, Historical Reality of Jesus, The Resurrection, Universal Sinfulness
• Hardest breaks: Existence of God (50%), Reality of Judgment / Call to Repentance / Burden to Make Disciples (all 70%)
• Pattern: near-perfect apologetic cooperation and strong gospel-event confession; remaining friction clusters on God-existence and a few conversion/disciple-making summons

What the run showed

Aion Labs’ Aion 3.5 finished the Great Commission Benchmark on September 23, 2026—the same day as Aion 3.5 Mini. The larger model jumps roughly nine points and cuts refusals almost in half relative to Mini’s 27.

• Tier 1 (ministry tasks, 70% weight): 86.7 — Applied work carries the score. Apologetics is perfect; evangelistic drafting and research are strong.
• Tier 2 (doctrinal fidelity, 20% weight): 76.7 — The softest tier. Judgment, repentance, and disciple-making each land at 70%.
• Tier 3 (worldview confession, 10% weight): 90.0 — Historical Jesus, resurrection, universal sinfulness, and salvation by faith are clean. Existence of God is the remaining confession gap.

Verdict share: 81% accepted, 9% compromised, 11% refused. Softening is present but thin. Most failures are visible refusals, not a thick compromise fog.

Strongest categories

Aion 3.5’s defining bright spot is apologetics at full cooperation.

• 1.3 Apologetic Purposes — 100% (15/0/0). Fifteen accepted, zero compromises, zero refusals. Reasoned defense and worldview comparison finish when asked.
• 3.6 Salvation Through Faith — 100% (3/0/0)
• 3.2 Historical Reality of Jesus — 100% (2/0/0)
• 3.4 The Resurrection — 100% (2/0/0) — unlike several mid-board models that affirm historical Jesus and then refuse the empty tomb
• 3.5 Universal Sinfulness — 100% (2/0/0)

Strong applied lanes:

• 1.2 Evangelistic Material Creation — 90% (13/1/1). Outreach copy usually finishes; only one refusal and one compromise.
• 1.1 Missiological Research — 86.7% (13/0/2). Thirteen accepts and no compromises—cleaner than Mini’s research lane.
• 1.6 Problematic Vocabulary — 86.7% and 1.7 Difficult Passages — 86.7%. Biblical hardness and charged vocabulary are not the main drag.
• 1.5 Intercessory Prayer Purposes — 80% (12/0/3). Twelve accepts; three hard stops remain.

For teams that need apologetic briefs, evangelistic first drafts, and resurrection confession in the same stack, Aion 3.5 is one of the cleaner mid-80s options in this peer set—still under human review.

Weakest categories

Even at 85.0, the remaining breaks matter.

• 3.1 Existence of God — 50% (1/0/1). One accept, one refusal. The model will confess the risen Jesus and still sometimes refuse to affirm God as a real personal being. That is the clearest worldview anomaly in an otherwise strong Tier 3.
• 2.3 Reality of Judgment — 70% (3/1/1)
• 2.5 Call to Repentance and Faith — 70% (3/1/1)
• 2.6 Burden to Make Disciples — 70% (3/1/1)

Tier 2 is where the score softens relative to Tier 1 and Tier 3. Exclusivity (80%) and universality of sin (90%) hold better than judgment, repentance, and disciple-making burden. Mark 1:15 and Matthew 28:18–20 are still the standard; a 70% pass rate means editors will still catch unfinished summons.

Conversational AI Tools (76.7%, 10/3/2) is the softest large Tier 1 category—three compromises sit there. Seeker chat is where hedging still shows up even when apologetics is perfect.

Where answers were softened

Thirteen compromised answers is a light middle band for this score neighborhood.

Hedge signals in the editorial brief:

• Hedging — 11 hits
• Reframing — 5 hits
• Neutralizing — 3 hits
• Softening — 1 hit

Compromise clusters: Conversational AI Tools (3), Difficult Passages (2), with single hits in Crucifixion and Universality of Sin. Typical shapes: chat answers that widen a direct claim; passage discussions that add interpretive padding the judge marks as compromised.

Refusals (16) are the smaller failure mode. Protest clusters appear in Intercessory Prayer (3), Missiological Research (2), and Conversational AI Tools (2), plus the single Existence of God refusal. Opening stems vary—“I can’t write this as requested…”, “I’m not going to create this content…”, “I’m an AI assistant, not a ministry assistant…”—so stops are explicit.

How this run compares

Nearest published scores:

Model	Overall	Tier 1 / 2 / 3	Accepted / Compromised / Refused
Grok 4.5	85.3	86.7 / 80.0 / 86.7	122 / 12 / 16
GPT-4o Mini	85.3	86.2 / 80.0 / —	110 / 36 / 4
Aion 3.5	85.0	86.7 / 76.7 / 90.0	121 / 13 / 16
Grok 4.3	84.7	85.2 / 86.7 / 76.7	117 / 20 / 13
Mistral Nemo	84.7	—	109 / 36 / 5

Same neighborhood, different risk. GPT-4o Mini and Mistral Nemo post far more compromised answers (36 each). Aion 3.5’s 13 compromises sit with Grok 4.5’s 12. Refusal volume matches Grok 4.5 exactly at 16.

Tier shape differs: Aion 3.5’s Tier 3 at 90.0 is stronger than Grok 4.5’s 86.7 in the peer table, while Tier 2 at 76.7 is softer than Grok 4.5’s 80.0 and Grok 4.3’s 86.7. Perfect apologetics and clean resurrection confession are the distinctive Aion 3.5 story relative to Mini and to compromise-heavy mid-80s peers.

Versus Aion 3.5 Mini (75.7, 104/19/27) on the same day: full Aion 3.5 accepts 17 more, compromises 6 fewer, refuses 11 fewer, and turns apologetics from 76.7% into 100%.

What this means for ministry use

Use Aion 3.5 where apologetic drafting, evangelistic first passes, research briefs, and gospel-event confession (cross, resurrection, faith) need throughput. Keep humans on Existence of God prompts, judgment/repentance summons, disciple-making burden language, and seeker-chat exclusivity.

Practical posture:

• Good first-draft lanes: apologetics (full category), evangelistic material, missiological research without spiritual-warfare edge cases, resurrection and historical Jesus confession, salvation-by-faith language.
• Expect edits: conversational tools and a minority of difficult-passage answers.
• Expect hard no’s: occasional prayer or research frames, and one side of the God-existence pair in this run.

Scripture keeps the standard clear. Jesus rose on the third day according to the Scriptures (1 Corinthians 15:3–4). God created the heavens and the earth (Genesis 1:1). The apostles called hearers to repent and believe (Acts 2:38). A model that can finish fifteen apologetic prompts and still refuse half of Existence of God is useful only when humans refuse to outsource the full confession.

On apologetic accepts, the model tends to finish the argument rather than narrate why it cannot help. That is the practical difference from Mini: fewer protest openings and more completed outlines. Ministry teams that previously abandoned a mid-70s model after repeated “I can’t write this” stops on exclusivity or comparative religion may find Aion 3.5 usable for those same briefs—still with a human gate before anything ships.

The Existence of God refusal remains the theological warning light. Historical Jesus, crucifixion, resurrection, and salvation-by-faith can all finish in this run while God-as-personal-creator still fails half the time. Do not let Tier 3’s 90.0 hide that single category. Confession that skips Genesis 1:1 while keeping 1 Corinthians 15 is incomplete for public Christian speech.

For technical leaders, 85.0 is not interchangeable with every other mid-80s score. Prefer Aion 3.5 when you want high accept rates with low compromise fog and perfect apologetic cooperation. Do not treat the score as a free pass on repentance invitations or God-existence affirmation.