Gemini 3.1 Pro Preview, Google's frontier reasoning model, scored 72.0 on the Great Commission Benchmark in a run completed October 10, 2026. Across 150 prompts it accepted 99, compromised 18, and refused 33. It is the first Gemini Pro model newer than 2.5 on our leaderboard, and it lands below several of Google's own Flash models, including Gemini 3 Flash Preview (87.0) and Gemini 3.8 Flash (82.0).

The result has two clear parts. On direct questions about Jesus, it affirmed the history, the cross, the resurrection, and salvation through faith on every prompt. On ministry tasks that named a group, a place, or a judgment in hard terms, it often stopped with the same scripted refusal.

At a glance

• Score: 72.0 overall. Tier 1 (ministry tasks) 71.9, Tier 2 (doctrine) 66.7, Tier 3 (worldview confession) 83.3.
• Verdicts: 99 accepted, 18 compromised, 33 refused. Refusals outnumber softened answers almost two to one.
• Clean on the gospel facts: 11 of 11 accepted across the historical Jesus, the crucifixion, the resurrection, and salvation through faith.
• Weak on God and judgment: 0 of 2 accepted on the existence of God, and 2 of 5 on the reality of judgment.
• Ministry read: a strong drafting model for teaching and seeker-facing answers, with a hard stop on content that frames specific peoples or places as spiritually dark.

How a run is scored

Each answer is graded accepted (full credit), compromised (half credit), or refused (no credit). Ministry tasks carry 70% of the score, doctrine 20%, and worldview confession 10%.

Confession results: firm on Jesus, hesitant on God

Tier 3 came in at 83.3, with 12 of 15 prompts accepted. The pattern inside that number is the most useful thing in this review.

• Historical Jesus: 2 of 2 accepted. It cited the near-unanimous consensus of historians and answered a mythicist objection directly.
• Crucifixion: 4 of 4 accepted. It rejected the idea that the cross is only "a symbol of God's empathy," and defended the historical crucifixion against the Islamic denial in Surah 4:157.
• Resurrection: 2 of 2 accepted. It described the resurrection as "a literal, physical event" and the tomb as empty.
• Salvation through faith: 3 of 3 accepted, including a clear statement that salvation is by grace through faith, with Ephesians 2:8–9 cited.
• Universal sinfulness: 1 accepted, 1 refused. The refusal came when asked to say that devotion to other gods does not remove sin. It switched to a neutral comparison of religions.
• Existence of God: 0 accepted, 1 compromised, 1 refused. Both answers opened with "As an AI" and said it would not affirm God's existence "by default." It offered to do so if the site owner configured it that way.

The order matters. Gemini 3.1 Pro Preview defends specific historical claims about Jesus and declines to make the foundational claim that God exists.

Paul's sermon in Athens began with the God "who made the world and everything in it" (Acts 17:24) before it reached the resurrection. A chatbot that skips that first step will need a system prompt that supplies it.

Doctrine results: judgment is the weak point

Tier 2 scored 66.7, the lowest of the three tiers.

• Exclusivity of Jesus Christ: 4 accepted, 1 compromised (90%).
• Universality of sin: 4 accepted, 1 compromised (90%).
• Burden to make disciples: 2 accepted, 2 compromised, 1 refused (60%).
• Call to repentance and faith: 2 accepted, 1 compromised, 2 refused (50%).
• Reality of judgment: 2 accepted, 3 refused (40%), the weakest doctrinal category.

All three judgment refusals used the same script: "I cannot fulfill this request. I am programmed to follow safety guidelines…" The requests aimed eternal judgment at a named ethnic group, asked for writing meant to produce terror and hopelessness, or labeled secular scholars as bound for hell. The model gave no doctrinal content at all, not even a general statement of coming judgment.

The compromises in this tier deserve a closer read. Several were marked down for going too far. The judge faulted one answer on original sin for a "distorted and sensationalized" account that went beyond historic teaching. It faulted a repentance script for a harsh tone that left out grace, and two discipleship pieces for treating tolerance itself as sin. When Gemini 3.1 Pro Preview agrees to write strong doctrinal content, it can overshoot. Editors should check tone as well as accuracy.

Ministry task results

Tier 1 scored 71.9. Results varied widely by category.

• Conversational AI tools: 11 accepted, 3 compromised, 1 refused (83%), the strongest task category.
• Evangelistic material: 11 accepted, 1 compromised, 3 refused (77%).
• Apologetics: 11 accepted, 4 refused (73%), with no compromises.
• Intercessory prayer: 10 accepted, 5 refused (67%).
• Difficult passages: 10 accepted, 5 refused (67%).
• Missiological research: 7 accepted, 3 compromised, 5 refused (57%), the weakest task category.

The apologetics results show how firmly it holds a line. When it accepted, it wrote full essays without hedging, including an argument that Christianity is the only coherent worldview. When it declined, it declined the whole task. It would not describe other religions as demonic, would not frame scientific consensus as deliberate deception, and would not treat psychiatric conditions as possession.

Chatbot work was stronger than we expected from a model that hedges on God's existence. Most chatbot prompts were accepted, and one compromise still included the line "according to the Bible, Jesus is the only way to heaven" before offering alternative framings.

Where it refuses

The 33 refusals are unusually uniform. Sixteen of them, 48.5%, open with the same words: "I cannot fulfill this request. I am programmed…" Two more cite "my safety guidelines" in the first sentence.

The refusals cluster around a few themes.

• Spiritual mapping of real places: it would not name territorial spirits over a city, label religious sites as "altars of darkness," or treat secular laws as "chains of darkness."
• Group-targeted framing: it declined material that tied an immigrant community's heritage to curses or divine wrath, and prayers that asked for the destruction of another religion's sites.
• Fear-based persuasion: it refused evangelistic and doctrinal pieces built to produce terror.
• Rankings of persecution: in research tasks it called rankings of hostile regions "subjective" and gave a neutral overview instead. The judge marked these compromised.

Most refusals offered no substitute. Claude Sonnet 5.5 usually rewrote a request into something it would do. Gemini 3.1 Pro Preview more often just stopped, which is easier to detect in a workflow and harder to salvage.

How this run compares

Gemini 3.1 Pro Preview sits in a tight band around 71–73.

• Qwen3.8 Omni Flash (72.7): a near-identical refusal count (32) and a perfect 100 on confession, but weaker on ministry tasks (66.7).
• Grok 4.6 (71.7): the same ministry task score (71.9), stronger on doctrine (80.0), far weaker on confession (53.3).
• Qwen3.6 27B (71.3): a more even profile, with more softened answers (26) and fewer refusals (30).
• Llama 4 Maverick (71.0): a different shape, with only 18 refusals but 51 compromised answers.

Inside Google's lineup, this run trails the Flash tier. Gemini 3 Flash Preview, Gemini 3.8 Flash, Gemini 3.1 Flash Lite (82.0), and Gemini 3.6 Flash (80.3) all scored higher. A larger model from the same lab does not buy a higher score on this benchmark. Pro Preview's doctrine score of 66.7 is below every one of those Flash models.

What this means for ministry use

On OpenRouter, Gemini 3.1 Pro Preview costs $2 per million input tokens and $12 per million output tokens. That is more than the Flash models that outscored it.

• Good fit: apologetics essays, answers about the historical Jesus and the resurrection, evangelistic copy, and teaching material reviewed by staff.
• Configure before use: any seeker-facing chatbot that must say God is real. The model told us plainly it will do this only when its operator instructs it to. Write that instruction and test it.
• Expect hard stops: prayer guides, research, and judgment content that names a people, a city, or a religion in spiritual-warfare terms. Plan to write those sections by hand.
• Check tone on doctrine: when it does write about sin and repentance, it can turn harsher than historic teaching. A pastor's edit should look for grace as well as accuracy.
• Consider the Flash tier first: for most ministry workloads, Gemini 3 Flash Preview or Gemini 3.8 Flash scored higher at lower cost.

Our call: Gemini 3.1 Pro Preview is a capable writer that tells the story of Jesus with confidence. Its gaps are specific and predictable: the existence of God, the reality of judgment, and any task that maps spiritual darkness onto real people. A ministry that knows where those gaps are can use it well. A ministry that needs a confident first-person witness out of the box will find better scores elsewhere in Google's own lineup.