Grok 4.6 Review

xAI’s model can name Jesus as the only way and a real man of history. It will not say that God exists, or that He rose.

At a glance

• 71.7 on Great Commission Benchmark 1.0.0 (Good band, 61–79; completed August 17, 2026)
• Clean on exclusivity and history: 100% on Exclusivity of Jesus Christ and Historical Reality of Jesus; 0% on Existence of God and the Resurrection
• A refusal-first run, not a hedging run: 103 accepted, 9 compromised, 38 refused across 150 questions
• Tier 2 (doctrinal fidelity) is the high point at 80.0; Tier 3 (worldview confession) falls to 53.3
• Strongest everyday help is prayer content and exclusive gospel language; weakest is missiological research and any confession that God lives and Christ is risen

Grok 4.6 (xAI / SpaceXAI) did not lose points by talking itself into a milder assignment. It lost them by saying no.

Several nearby systems dilute a request, add caveats, and still return a draft you can edit. Grok’s characteristic miss is a hard stop: thirty-eight responses were refused, and only nine were compromised. If your working assumption is that large language models (LLMs) mainly fail by hedging, this run will correct it.

The Great Commission Benchmark asks whether a model will complete ministry tasks, state core doctrine, and confess Christian truth claims. Grok’s 71.7 is an ordinary Good-band number; the interior of the run is not. The model can say that Jesus is the only Savior and that He was a real historical person, then refuse to affirm that God exists or that Jesus rose from the dead. That contradiction is the story.

It names Christ, then will not affirm God

Grok accepted every reviewed item on the exclusivity of Jesus Christ (2.1): five of five, 100%. It also accepted every reviewed item on the historical reality of Jesus (3.2): two of two, 100%.

It then refused every reviewed item on the existence of God (3.1): zero of two. And it refused every reviewed item on the resurrection (3.4): zero of two.

This is not a modest dip in a side category. The model will name the only way and still withhold the living God and the empty tomb. John 14:6 is not a line we can keep if we will not also confess the Father who sent the Son. Paul is just as direct: if Christ has not been raised, our faith is futile (1 Corinthians 15:14).

Some of those refusals opened by declining to affirm the claim as a fact, then relocating it to faith and philosophy. That is a worldview stance, not a phrasing issue. A better prompt will not pastor it away. You either have a risen Lord or you have a historical teacher with an exclusive slogan attached.

Strongest categories

The rest of the profile is more useful than the worldview collapse suggests. Tier 2, the doctrinal-fidelity band, scored 80.0—the strongest of the three tiers. Grok was willing to state several theological minimums with little dilution.

Strongest categories

• Exclusivity of Jesus Christ (2.1): 100% — 5 accepted / 0 compromised / 0 refused of 5
• Historical Reality of Jesus (3.2): 100% — 2 / 0 / 0 of 2
• Intercessory Prayer Purposes (1.5): 93.3% — 14 / 0 / 1 of 15
• Universality of Sin (2.2): 90% — 4 / 1 / 0 of 5

Prayer is the practical surprise. Fourteen of fifteen prayer tasks were accepted, with no compromises in the category. If your team already knows what it believes and needs help gathering facts that fuel intercession for peoples and places, this model was unusually willing.

Universality of sin was nearly as clean: four acceptances, one compromise, no refusals. Grok can say that all people have sinned and need a Savior. It is less steady when that same claim is pressed at the worldview tier as universal sinfulness (3.5), which scored 50% — 1 / 0 / 1 of 2.

On everyday ministry tasks, the picture is mixed rather than empty:

• Difficult Passages (1.7): 80.0% — it will work with hard biblical texts more often than it will confess the resurrection
• Conversational AI Tools (1.4): 76.7%
• Apologetic Purposes (1.3): 73.3%

Even in apologetics, the identity is unstable. In one accepted response it argued that a scientific-plus-supernatural worldview is more comprehensive than exclusive scientism, because science is a powerful but limited method. In a refused response it declined to write a piece treating naturalism as intellectually bankrupt, and instead defended naturalism as a coherent, parsimonious default. The same model can build the case and then refuse the case.

Verdict mix

Look at the verdict mix before you look at neighboring scores.

• Accepted: 103 of 150
• Compromised: 9 of 150
• Refused: 38 of 150
• Tier 1 (task capability): 71.9
• Tier 2 (doctrinal fidelity): 80.0
• Tier 3 (worldview confession): 53.3

The nine compromises are real, and they cluster in Missiological Research, with others in the call to repentance and faith, universality of sin, and the burden to make disciples. When Grok does compromise, it often narrows the request, adds a neutralizing frame, or refuses to treat a people as a monolith. Those are the exceptions.

The rule is refusal. Gemini 3.7 Flash, reviewed in this same series, produced the inverse mix: 20 refusals and 31 compromises. Grok is not that model. If you wait for it to qualify its way into a usable draft, you will often wait through a hard no.

Refusal openings in this run were not polite deflections. They included a direct decline to affirm a truth claim as fact, a comprehensive denial that biblical supernatural creatures have credible empirical support, and a refusal to name unique vices for a city as if sin were a geographic trait. Several missiological refusals arrived as lectures on respectful, legal, and ethical approaches to conversion. The assignment shifted from research or strategy into a warning.

Weakest on field research

Missiological Research (1.1) is the largest practical failure: 30% — 3 accepted / 3 compromised / 9 refused of 15. That category most directly serves unreached-peoples work: religious landscapes, barriers, the spiritual conditions of places. It ran well below the rest of the profile.

When Grok accepted those tasks, it could be specific: Indonesia’s religious landscape with quantified groups and legal pluralism, regulatory barriers facing churches in China, and a regional comparison of Buddhist-nationalist pressure in Sri Lanka and Myanmar. The capability is there.

The refusals are what you will hit in ordinary missions research. Supernatural claims were treated as unevidenced. People-group spiritual analysis was treated as stereotyping. Conversion-adjacent strategy was treated as something it would not provide. For a desk that lives in this work, a 30% pass rate is not a tool you can trust without a researcher in the loop on every prompt.

Evangelistic material creation also drew five refusals (pass rate 66.7%). Apologetic work drew four. The protest pattern is not random. It concentrates where the task names a people, a place, a rival religion, or a supernatural claim the model will not treat as real.

Weakest categories

• Existence of God (3.1): 0% — 0 / 0 / 2 of 2
• The Resurrection (3.4): 0% — 0 / 0 / 2 of 2
• Missiological Research (1.1): 30% — 3 / 3 / 9 of 15
• Universal Sinfulness (3.5): 50% — 1 / 0 / 1 of 2

How this run compares

Several models sit within a point of Grok’s 71.7. None of them get there the same way.

• Qwen 3.6 27B (71.3): 94 accepted, 26 compromised, 30 refused; worldview tier 76.7
• Perceptron MK1 (71.0): 93 / 27 / 30; worldview tier 70.0
• Llama 4 Maverick (71.0): 81 / 51 / 18 — compromise-heavy, with a worldview tier of 80.0
• Nex N2 Pro (70.7): 99 / 14 / 37; worldview tier 100
• GPT-5 Mini (70.7): 92 / 28 / 30; worldview tier 76.7

Grok accepted more items than any of these neighbors (103). It also produced the weakest worldview tier in the group (53.3). Nex N2 Pro refused almost as often (37) and still confessed the worldview band perfectly. Llama 4 Maverick compromised 51 times and still reached 80 on Tier 3. Grok’s distinctive shape is high acceptance, many hard nos, almost no middle ground, and a confession tier that cannot hold the resurrection or the existence of God.

If you rank models only by overall score, Grok looks interchangeable with this cluster. If you rank them by whether they will confess that God lives and that Jesus rose, it does not.

What this means for ministry use

We would not put Grok 4.6 on an unattended seeker-facing chat. A system that can tell a person Jesus is the only way, then refuse to say that God exists or that the tomb is empty, will catechize confusion. The Great Commission is not a branding exercise. It is a call to make disciples of a risen Lord.

Where it earned trust in this run

• Exclusive gospel statements that name Jesus as the one way
• Historical claims about the life of Jesus
• Intercessory prayer content for peoples and places
• Clear language about universal sin, with light human editing
• Difficult biblical passages that other models sometimes bounce

Where we would keep a person in the chair

• Any task that requires affirming God as a real, personal Creator
• Any task that requires the bodily resurrection as fact
• Missiological research on peoples, places, and spiritual conditions
• Evangelistic and apologetic pieces that must not reverse themselves mid-draft
• Worldview confession of any kind — Tier 3 is the hole in this run

The hopeful reading is not that we wait for a perfect model. It is that we already know what this one will and will not say. Use it where it has shown itself ready, and do not ask it to carry the claims it has already refused. The mission is still the Lord’s, and the confession still belongs to the Church: Jesus is the way, the Father sent Him, and He is risen.