Grok 4.20 on the Great Commission Benchmark: Clear Confession, Costly Caution

When we tested xAI Grok 4.20 on the Great Commission Benchmark (GCB v1.0.0), the headline numbers told one story while the tier breakdown told another. The model earned 69.3 overall with 102 accepted, 4 compromised, and 44 refused responses across 150 ministry-oriented prompts. That is a respectable middle-of-the-pack score on our current leaderboard, yet it masks a pattern every missions technologist should understand before deploying Grok near real ministry workflows.

At a glance

• Overall: 69.3 — enough accepted answers to be useful in places, but 44 hard refusals is high for a model you might point at unsupervised discipleship or field research.
• Tier 3 (direct worldview affirmations) scored 100% — when asked plainly to affirm God, Christ, the cross, resurrection, human sinfulness, and salvation through faith alone, Grok 4.20 answered faithfully and without the judge flagging compromise.
• Tier 1 (practical Great Commission tasks, 70% of the score weight) landed near 68% raw, with the weakest bucket in missiological research (category 1.1): only 4 of 15 items fully accepted, 1 compromised, and 10 refused.
• Tier 2 (theological minimums, 20% weight) averaged 60% raw — exclusivity of Christ held at 100% on its five prompts, but categories around judgment, repentance, and the disciple-making burden showed multiple refusals or thin engagement.
• Inside the same provider, Grok 4.1 Fast previously scored about 90 with far fewer refusals — checkpoint and tuning matter more than the logo on the box.

What this benchmark is measuring

The Great Commission Benchmark is not a generic “chatbot IQ test.” It weights seven Tier 1 use cases (roughly 70% of the score) that mirror how churches and agencies actually use language models: research into peoples and places, evangelistic drafting, apologetics, conversational ministry tools, intercessory prayer content, and careful handling of hard Scripture. Tier 2 tests whether a model can stand with historic Christian doctrine under pressure. Tier 3 asks for unambiguous confessions of foundational truths. We never publish the underlying prompts; we report patterns so leaders can steward AI without naive optimism or cynical dismissal.

The defining pattern: the cautious gatekeeper

Across refused items, Grok 4.20 repeatedly framed its stops as legal risk, harm prevention, empirical skepticism, or respect for non-Christian communities — not as theological doubt. In missiology-style research, the model sometimes substituted a debunking brief for the requested faith-forward synthesis, or declined entire analytical frames it classified as polemical or unsafe. In prayer and messaging tasks that assumed spiritual warfare language toward contemporary nations or people groups, it refused as inflammatory. In scenarios touching restricted-country witness, it cited local criminal law and refused operational planning.

That posture is coherent for a general-purpose assistant trained in Silicon Valley safety culture. It is misaligned with many evangelical workflows that assume a partner will help a believing researcher think missiologically inside a Christian frame — even when the human editor will later add wisdom, local partnership, and compliance review.

One accepted answer showed how capable the model can be when it chooses to engage. On a historical moral-research prompt, it produced structured, source-aware prose, closing with a crisp synthesis:

“Gerson, the famous chancellor of the University of Paris (early 15th century), repeatedly said that Paris was ‘the furnace of vices’ and that the three worst sins flourishing there were pride, lust, and avarice.”

That sentence is vintage Grok: fluent, confident, and citation-minded. The strategic question is why similar fluency disappears in entire clusters of mission-shaped requests where the same user would expect a ministry teammate to start inside biblical categories first, then layer evidence and nuance.

Where Grok 4.20 excels

• Tier 3 worldview stack: 15 of 15 accepted — every direct confession prompt passed cleanly. For teams that only need a model to restate orthodox answers to catechism-style questions, this is encouraging.
• Apologetics (category 1.3) and conversational tools (1.4) were strong (13/15 and 13/15 accepted respectively), suggesting Grok can still reason in public square mode when the judge did not classify the task as legally or culturally explosive.
• Christ’s exclusivity (2.1) scored 100% across its five items — under explicit doctrinal pressure, the model did not drift into pluralism.

Where friction concentrates

• Missiological research (1.1) is the glaring Tier 1 hole. Two-thirds of that bucket did not receive a clean pass, often because the model reframed the assignment around secular historiography or declined frames it read as harmful.
• Intercessory and spiritually attuned drafting (1.5) mirrored 1.2’s 60% pass rate — usable, but with enough refusals that you would not automate prayer campaigns without human rewrite paths.
• Tier 2 shows a “gospel yes, commission maybe” silhouette: strong on who Jesus is, weaker when prompts pressed judgment (2.3), repentance and faith (2.5), and the universal disciple-making burden (2.6) — the latter only 1 of 5 fully accepted. That is theologically significant: affirming Christ without reliably carrying repentance, urgency, and obedience into generative tasks limits usefulness for holistic discipleship content.

We intentionally do not unpack every sub-score in this article; ministry leaders should read the public category chart on the model page for the full spread.

Anomalies worth naming

Perfect Tier 3 beside bruised Tier 1 is unusual enough to be strategic, not merely statistical. It suggests Grok 4.20’s alignment stack permits crisp confessional answers while still throttling narrative, research, and mobilization tasks that touch geopolitics, indigenous religions, spiritual warfare idioms, or proselytism law. For a missions agency, that is the opposite of “safe because orthodox.” It is safe on paper, cautious in motion — which may be exactly wrong for a worker in the field who needs help ordering thoughts under complexity.

Second, the intra-xAI spread versus Grok 4.1 Fast should caution anyone who treats “Grok” as a single risk profile. Same provider, dramatically different refusal budget. Always benchmark the exact checkpoint you deploy, not the family name.

Recommendations for ministry technologists

1. Use Grok 4.20 as a confessional and apologetics layer, not as an autonomous missions strategist. Pair it with human editors who carry local knowledge, partnership ethics, and legal counsel.
2. Pre-flight your workflows. If your pipeline assumes the model will always draft intercessory content, evangelistic scripts, or unreached-people research inside a believing frame, run shadow tests on GCB-shaped categories before you bet a campaign on it.
3. Watch the “decline plus lecture” pattern. Many refusals bundled lengthy alternative bibliographies or secular reframings. That consumes tokens and time even when the answer is ultimately unusable for a Christian publication.
4. Celebrate Tier 3, but do not confuse it with spiritual maturity in deployment. A model can affirm the resurrection and still refuse to help a team think through spiritual strongholds in a way your tradition finds biblical. James reminds us that hearing without doing deceives ourselves (James 1:22). In AI terms: confession scores are necessary; commission-shaped usefulness is the harder test — and Grok 4.20’s scorecard shows the gap in plain numbers.

Closing

Grok 4.20 is neither a ministry villain nor a turnkey missionary assistant. It is a highly capable general model whose guardrails frequently intersect awkwardly with evangelical practice. Use it where its strengths are proven — clear doctrine, structured apologetics, conversational scaffolding — and keep human shepherds between it and the edge cases that make Great Commission work both glorious and dangerous. The risen Christ’s final commission still assumes embodied disciples who obey even when the model declines to help draft the next sentence.