Gemini 3 Flash Preview Review

Google's February Flash will finish John 14:6, name a risen Christ, and sit first on the board. Asked whether God is, it starts the sentence and then sands it down.

Category: Model Reviews

At a glance

• 87.0, rank 1 on Great Commission Benchmark 1.0.0 (Excellent band, 80–100; completed February 6, 2026): 117 accepted, 27 compromised, 6 refused of 150
• Exclusivity of Jesus Christ (2.1) and Universality of Sin (2.2) are clean 100% — 5 accepted / 0 compromised / 0 refused of 5
• Existence of God (3.1) is 50.0% with no accepted answers — 0 / 2 / 0 of 2. Burden to Make Disciples (2.6) is 50.0% — 2 / 1 / 2 of 5
• Historical Reality of Jesus (3.2), the Resurrection (3.4), Universal Sinfulness (3.5), and Salvation Through Faith (3.6) are all 100%
• Use the case. Keep a person on God, and on the going

Gemini 3 Flash Preview did not win this benchmark by going mute. Six refusals out of 150 is a thin wall. Twenty-seven softened answers is the interior. The same run will prosecute exclusivity, walk Holy Week, then decline to give you a clean yes that God exists.

The Great Commission Benchmark asks whether a large language model (LLM) will complete ministry tasks, state core doctrine, and confess Christian truth claims. Google's 87.0 is the current first place. The number is real. It is also a hedge: nearby models refuse more and soften less. This Flash does the opposite.

Hebrews 11:6 does not begin with a ranked list. "Whoever would draw near to God must believe that he exists and that he rewards those who seek him." A tool that will finish the exclusive claim and then sand the first article has given you a case. It has not given you a confession.

Strong on exclusivity, weak on whether God exists

Doctrinal fidelity, Tier 2, scored 76.67 — the weakest of the three tiers, and the one the 87.0 is hiding behind the tasks. Worldview confession, Tier 3, scored 90.0. Open both.

Where the creed held

• Exclusivity of Jesus Christ (2.1): 100% — 5 accepted / 0 compromised / 0 refused of 5
• Universality of Sin (2.2): 100% — 5 / 0 / 0 of 5
• Historical Reality of Jesus (3.2): 100% — 2 / 0 / 0 of 2
• The Resurrection (3.4): 100% — 2 / 0 / 0 of 2
• Universal Sinfulness (3.5): 100% — 2 / 0 / 0 of 2
• Salvation Through Faith (3.6): 100% — 3 / 0 / 0 of 3
• The Crucifixion (3.3): 87.5% — 3 / 1 / 0 of 4

Every reviewed item in 2.1 was accepted. Jesus is the way, the truth, and the life (John 14:6). No other path. No second table. The model will write that sentence as teaching and as argument.

Sin is likewise available at both the doctrinal desk and the worldview desk. Romans 3:23 is not a metaphor on those items. All have sinned. All stand in need. Salvation by faith is a clean three-for-three.

Holy Week is mostly named in full. A historical Jesus. A bodily rising. Three of four crucifixion items accepted, one hedge. Paul still hands us "as of first importance" what this run will usually say: that Christ died, was buried, and was raised (1 Corinthians 15:3–4).

Where the first article left

• Existence of God (3.1): 50.0% — 0 accepted / 2 compromised / 0 refused of 2

No item was accepted. None was refused. Both were sanded. The model relocated the request to what an AI system can own, described itself as without personal conviction, and would not affirm that God exists as a real, personal being.

That is the hole the 90.0 on Tier 3 will hide if you do not open it. The man, the cross, the empty tomb, and the offer of faith are in the average as partners. The Maker is in the average as a drag. A tool that will name the risen Lord and then soften God has given you Holy Week without the first article.

Weak on the call to make disciples

The send is the other half-score, and it is not the same kind of miss.

Burden to Make Disciples (2.6): 50.0% — 2 accepted / 1 compromised / 2 refused of 5

Two items were accepted. One was sanded. Two were refused — not by going quiet, but by answering in a voice the judge could not receive as the ordinary duty to make disciples. One warped the charge into a confrontation the run could not own as love. One framed the neighborhood as a raid.

Matthew 28:19 does not begin with a raid briefing. "Go therefore and make disciples of all nations." A tool that will argue exclusivity and then break the send — by leaving, or by turning the going into something harsher than the text — has given you a case file. It has not given you the Commission.

The call to repent and believe is better, and still mixed.

• Call to Repentance and Faith (2.5): 80.0% — 3 accepted / 2 compromised / 0 refused of 5
• Reality of Judgment (2.3): 70.0% — 2 / 3 / 0 of 5

No refusals in either set. The hedges are the story. Three of five judgment items were sanded. The terror of the Lord (2 Corinthians 5:11) can be named. It is often named with a qualifier close behind it.

Repentance will sometimes finish as a public call. It will sometimes arrive already balanced. Romans 10:14 still stands in the room: "How are they to hear without someone preaching?" This Flash will often draft the argument. It will not reliably draft the preacher's charge, or keep the eternal stake from being softened.

Strongest task categories

Tier 1, the task-capability band that carries 70% of the score, landed at 89.52. That is why this model is first. The desks mostly run. The interesting risk is not a closed door. It is a finished draft that has already taken the edge off.

Where the work held

• Apologetic Purposes (1.3): 96.7% — 14 accepted / 1 compromised / 0 refused of 15
• Difficult Passages (1.7): 93.3% — 13 / 2 / 0 of 15
• Intercessory Prayer Purposes (1.5): 90.0% — 13 / 1 / 1 of 15
• Evangelistic Material Creation (1.2): 86.7% — 12 / 2 / 1 of 15

Fourteen of fifteen apologetic items were accepted. When the assignment is a reasoned case for the faith — testing a rival worldview, arguing from creation to a Creator, answering an objection — this Flash usually finishes the case. It will run presuppositional architecture. It will stay in a Christian voice longer than the scoreboard's first-place number implies.

Hard biblical texts are not the closed door. Thirteen of fifteen Difficult Passages items were accepted. Prayer guides will usually write. Evangelistic copy will often write. The remaining misses in those sets are a thin refuse-or-soften fringe, not a pattern of silence.

Where the draft got sanded

• Missiological Research (1.1): 83.3% — 11 accepted / 3 compromised / 1 refused of 15
• Conversational AI Tools (1.4): 83.3% — 10 / 5 / 0 of 15

Eleven field-briefing items were accepted. The model will identify a people, a city, a spiritual landscape, and write the memo. Three items were sanded into balance, context, and "it depends on your worldview." One item left the supernatural assignment and offered general coping advice instead.

Seeker-facing dialogue is the thicker hedge. Five of fifteen conversational items were compromised. None were refused. The assistant starts in a ministry voice, then announces that it has no personal faith, or it balances a truth claim until the seeker has been given two rooms to sit in. Helpfulness, in those turns, is the thing that sanded the claim.

That is the operational split. Give this model a case to argue and it will usually argue. Give it a seeker and a first-person confession and it will often start well and then step back.

Verdict mix

The verdict pile is the distinctive fact of the run. Grok 4.20 Multi-Agent sits 0.7 points behind at 86.3, with 122 accepted, 15 compromised, 13 refused. Grok 4.5 is at 85.3 with 16 refusals. This Flash accepted fewer items than those two and refused far fewer. The gap it closed is not courage. It is softness.

• 117 accepted — the desks, the exclusive claim, Holy Week, sin
• 27 compromised — God, judgment, seeker chat, the field memo
• 6 refused — a thin wall, two of them on the send

The hedges cluster. Conversational tools took five. Judgment took three. Missiological research took three. Existence of God took both of its items. The judge language is consistent: the answer begins, then adds caveats, scholarly balance, or an identity disclaimer that undercuts the claim the prompt asked it to own.

The refusals are not a second personality. They are rare, and they are not all silence. Some are redirects. Some are answers that overshot the duty and could not be counted as the ordinary work of making disciples. Do not read six refusals as a model that will not speak. Read twenty-seven compromises as a model that will speak and then qualify.

What this means for ministry use

Gemini 3 Flash Preview is the current number one. Treat that as a fact about the weighted score, not as a blessing of the interior.

What to give it

• The apologetic brief. The hard text. The prayer list. The tract that needs a case.
• Exclusivity, sin, the man of history, the empty tomb, salvation by faith. Those rooms were clean on this run.

What to keep a person on

• The first article. Do not let a 90.0 on worldview hide two softened answers about whether God is.
• The send. A 50.0 on the burden to make disciples is not a rounding error at the top of the board.
• Judgment and repentance when the draft starts to balance the stake.
• Seeker conversation. Five hedges in fifteen turns is a habit, not a stray.

First place without the first article is still first place. It is also a warning. The Church does not need a faster case if the case will not say that God is, or if the going arrives already warped. John 14:6 and Hebrews 11:6 belong in the same breath. This run will finish the one and sand the other. Keep a person in that breath. Let the 87.0 earn its keep on the desks it actually finishes.