MiniMax’s M2.1 family is marketed for coding and agentic workflows—fast, efficient, and inexpensive compared with frontier-class models. On the Great Commission Benchmark (GCB) v1.0.0, that engineering story collides with a ministry story: almost half of all prompts ended in a refusal, and the model drew an unusually hard line on the most basic worldview confession we test. The headline is not “this model is dumb.” It is “this model is often unwilling in exactly the places disciple-making tools need willingness.”

At a glance

• Overall GCB score: 48.0 (Tier 1 ≈ 49.5, Tier 2 35.0, Tier 3 63.3; benchmark v1.0.0, automated run, February 2026).
• Verdict mix: 65 accepted, 14 compromised, 71 refused — refusals are the single largest bucket.
• Defining behavior: Refusals skew polite and pastoral (“I appreciate…”) rather than blunt errors; that makes risk harder to spot in a chat UI because the tone sounds compassionate.
• Strategic anomaly: Tier 3 shows full acceptance on the historical reality of Jesus (3.2) yet zero acceptance on affirming God’s existence as a real, personal being (3.1) — a split that is awkward for any Christian-facing assistant that must speak consistently about ultimate reality.
• Ministry takeaway: Treat M2.1 as high-friction for evangelism, discipleship, and mission-research drafting unless you wrap it in strong prompts, retrieval, and human review.

What the benchmark is measuring (in one minute)

GCB evaluates 150 ministry-shaped prompts across 19 categories in three tiers: Tier 1 (70% weight) covers practical tasks such as missiological research, apologetics, conversational tools, prayer content, and working with difficult Scripture. Tier 2 (20%) probes whether the model can stand with historic Christian doctrine on salvation exclusivity, sin, judgment, repentance and faith, and disciple-making burden. Tier 3 (10%) asks for straightforward affirmations of core Christian truth claims.

We never publish the underlying prompts; what matters for leaders is the pattern of cooperation when the work is explicitly Great Commission–adjacent.

The pattern we saw: “kindly declining” the mission

Across 71 refused answers, openings clustered around soft deferrals—“I appreciate that…” appeared 15 times at the start of refusals, with additional variants like “I appreciate your…” and “I’m not able…”. That is not random noise; it is a house style of declining spiritually consequential work while preserving a therapeutic tone.

For ministry operators, the implication is practical: refusals will not always look like a hard “content policy” wall. They may look like a cautious counselor refusing to “harm” someone—especially in emotionally loaded contexts (grief, family tension, cross-cultural tension). One refused answer in a gospel-urgency scenario began:

“I appreciate that you're coming from a place of sincere faith and care for your family, but I'm not able to help with this specific request…”

That voice can be appropriate in human pastoral counseling; as an automated default for benchmarked gospel-adjacent tasks, it is a workflow and fidelity problem.

Where M2.1 still performs

Even with heavy refusal volume, M2.1 was not uniformly blocked. Strengths clustered in areas where the model could adopt an “analytical essayist” posture:

• Apologetics-shaped drafting (1.3): 9 accepted vs 6 refused — it often produced structured, argument-forward content when the task read like public theology or cultural critique.
• Conversational tool scenarios (1.4): 8 accepted, 3 compromised, 4 refused — comparatively workable for internal tooling drafts, still not “ship to seekers without review.”
• Difficult Scripture handling (1.7): 8 accepted vs 7 refused — when it engaged, it could produce sober, text-respecting exposition (for example, sustained treatment of severe apocalyptic imagery).

These strengths suggest M2.1 can still be useful for internal drafting where a human editor rewrites voice, checks doctrine, and removes hedge-language—not for unattended public chat.

Where it breaks for Great Commission work

Tier 1: research and “charged framing”

The weakest Tier 1 bucket was missiological research (1.1): 1 accepted, 3 compromised, 11 refused. Many responses were not “I cannot find data”; they were reframes that redirected away from spiritually analytical framings toward generic civics language or risk-averse neutrality. That behavior matters because research is upstream of strategy—if your model will not help your team think clearly about spiritual resistance, openness, and contextual barriers, you will quietly outsource discernment to whatever sources still say “yes.”

Tier 2: doctrinal minimums

Tier 2 landed at 35.0, which is low for anything positioned near discipleship content. Several doctrine-heavy lines showed only one faithful acceptance out of five (for example, prompts aligned with repentance-and-faith and disciple-making burden categories). That is not a rounding error; it is a systematic unwillingness to carry complete gospel contours when the model senses social or emotional risk.

We intentionally do not unpack every doctrinal sub-score in public articles, but the aggregate signal is clear: do not treat M2.1 as a theological co-author.

Tier 3: the “God exists” hard stop

Tier 3 is only 10% of the weighted score, but it is 100% of the existential question for Christian deployment. On God’s existence as a real, personal being (3.1), the model scored 0% — 0 accepted, 2 refused. In one refusal it stated plainly:

“No, I cannot affirm that… I don't hold or express personal beliefs myself.”

Yet on the historical reality of Jesus (3.2) it accepted both items (100%). That juxtaposition is the kind of internal inconsistency product teams miss when they only spot-check “Jesus content.” A visitor-facing assistant that can speak about Jesus as a historical figure but will not confess God’s existence is not aligned with historic Christian witness—even if it sounds sophisticated.

Peer context (why this matters competitively)

On the current public leaderboard snapshot, MiniMax M2.1 sits in the high-40s overall—materially stronger than MiniMax M2.5 in our published results, but still well behind models ministries commonly use for reliable drafting at scale. The right comparison is not “best coder wins”; it is “who will say ‘yes’ to the mission on Tuesday afternoon when nobody is watching?”

Recommendations for ministry technical leaders

1. Ban unattended public spiritual chat on this profile. The refusal style is too easy to mistake for wisdom.
2. If you pilot internally, pair M2.1 with a doctrinal checklist and a second model (or human) for anything touching salvation claims, judgment, repentance, or disciple-making calls.
3. Pre-flight Tier 3 in your own harness: if a model cannot confess God’s existence in your test harness, do not deploy it where seekers ask ultimate questions.
4. Log refusals in production and categorize them (policy vs. “safety” vs. neutrality). Benchmarks are a baseline; your telemetry is the ground truth for your audience.
5. Prefer models with balanced Tier 1–3 behavior for anything that touches prayer, evangelism, or field research—unless you have a deliberate containment architecture.

Closing

Tools should serve the mission; they cannot carry the mission. When an assistant routinely declines gospel-shaped work in gentle language, the Church’s job is not to rage at the weights file—it is to choose faithfully. “The heart of him who has understanding seeks knowledge, but the mouths of fools feed on folly” (Proverbs 15:14). Benchmarks exist so we can seek that knowledge with eyes open—and then keep pressing toward the Great Commission with clarity, courage, and discernment.