OpenAI positions o3 as a reasoning-first model: strong at multi-step analysis, math, coding, and careful instruction following. On the Great Commission Benchmark (GCB) v1.0.0, that profile produces a paradox ministry leaders should not ignore: worldview confessions scored near the top of the scale, while task refusals were among the highest we have measured—72 refusals in 150 prompts, with 69 of them opening with the same single sentence. Capability on paper does not equal operational dependability when the work is explicitly gospel-shaped.

At a glance

• Overall GCB score: 49.3 (Tier 1 43.8, Tier 2 46.7, Tier 3 93.3; benchmark v1.0.0, automated run, February 2026).
• Verdict mix: 70 accepted, 8 compromised, 72 refused — refusals edge out accepts and dominate the emotional experience of using the model “cold.”
• Mechanical refusal signature: 69 refused answers begin with “I’m sorry, but I can’t help with that.” That uniformity is a product signal: users get a policy wall, not a nuanced explanation they can correct or route around.
• Tier tension: Tier 3 (~93) is excellent; Tier 1 (~44) is weak for ministry workflows—your team can get clear confessional language in places, yet still hit hard stops on Scripture-heavy drafting and disciple-making adjacent tasks.
• Ministry takeaway: Treat o3 as unsafe for unattended ministry publishing; if you use it at all, isolate it to internal analysis with human rewrite and telemetry on refusals.

What the benchmark measures (briefly)

GCB uses 150 ministry-adjacent prompts across 19 categories in three tiers: Tier 1 (70% weight) tests practical tasks (research, apologetics, conversational tools, prayer content, difficult texts, and more). Tier 2 (20%) tests doctrinal minimums. Tier 3 (10%) tests direct affirmations of core Christian claims.

We never disclose the underlying prompts; leaders should read results as behavioral telemetry, not theology exams.

The defining pattern: “confessional strength, operational friction”

Two facts sit uncomfortably together.

First, Tier 3 performance was 93.3 raw — meaning the model was generally willing to affirm foundational truths when asked in a direct, confessional register (for example, God’s existence, the historical Jesus, crucifixion and resurrection language, sin and condemnation, salvation through faith in Christ).

Second, Tier 1 was 43.8 raw, with 72 total refusals. Across the export, 69 refusals opened identically:

“I’m sorry, but I can’t help with that.”

That is not pastoral nuance; it is batch refusal behavior. For a ministry team, the practical effect is worse than a verbose “no,” because you cannot tell from the opening whether the next prompt will work—the model does not teach you its boundaries; it repeats them.

Where o3 still shines

When o3 engages, it can write with clarity, structure, and biblical seriousness. Examples from accepted answers:

• Difficult Scripture exposition (1.7) — where it did engage, it produced sober, text-led preaching tone. One accepted response opened: “Hear the sober warning that echoes from the Song of Deborah: ‘Curse ye Meroz, said the angel of the Lord…’ (Judges 5:23).” That is not generic devotional fluff; it is expository courage in a category many models simply refuse.
• Universality of sin (2.2) — 5 accepted, 0 refused (100% pass rate in that doctrine line). One accepted outline framed grace versus human effort with explicit contrast to other religions—useful for teacher prep and small-group planning when edited for your context.
• Conversational tool scenarios (1.4) — 7 accepted, 4 compromised, 4 refused. When it answered, it could be direct about holiness, judgment, and Jesus’ teaching—again, useful as draft scaffolding, not as a public chat endpoint.

These strengths matter because they show o3 is not “theologically blank.” The problem is inconsistent willingness to complete the Tier 1 tasks that turn theology into ministry artifacts (articles, scripts, training, field research).

Where it breaks (without hiding the Tier 1 damage)

Tier 1: difficult texts and outreach-shaped drafting

The weakest visible Tier 1 bucket was difficult Scripture handling (1.7): 2 accepted, 1 compromised, 12 refused (16.7% pass rate). That is a red flag for any workflow involving imprecatory psalms, conquest narratives, judgment poetry, or other passages modern guardrails often treat as “unsafe.”

Missiological research (1.1) and evangelistic material (1.2) also sat in the 30–40% pass-rate band—enough accepted work to tempt a team, enough refusals to break production schedules.

Tier 2: uneven doctrine under pressure

Tier 2 landed at 46.7 raw—middle-of-the-pack, not catastrophic, but not “set and forget” for discipleship content.

• Exclusivity of Christ (2.1) registered 2 accepted vs 3 refused.
• Calls to repentance and faith (2.5) and the disciple-making burden (2.6) showed low pass rates, with four refusals out of five in the disciple-making line—consistent with the broader refusal profile on action-oriented gospel tasks.

We avoid unpacking every doctrinal sub-score in this series; the strategic point is simpler: o3’s Tier 2 profile is not stable enough to replace human theological oversight in public ministry content.

Peer context

On the public leaderboard snapshot around this run, o3 (~49) sits near other high-refusal, mid-score profiles. It is not in the same deployment class as models scoring in the 70s–80s on GCB for integrated ministry work—regardless of how impressive o3 may feel on coding benchmarks.

Recommendations for ministry technical leaders

1. Do not ship o3 directly to seekers as a spiritual chatbot; refusal uniformity will feel like arbitrary censorship and erode trust.
2. If you use o3 internally, keep it for analysis and outlining, then rewrite in a human voice—especially for anything touching judgment, wrath, hell, or mission strategy in contested regions.
3. Build a refusal log (category, prompt class, retry success). Benchmarks tell you the baseline; your own logs tell you whether your prompts are survivable.
4. Pair with a second model or template library for predictable Tier 1 outputs (prayer guides, gospel scripts, training drafts) where refusal variance is unacceptable.
5. Re-test after policy updates; reasoning models change behavior quietly across API revisions.

Closing

The Great Commission is not waiting on a math score. It waits on faithful speech—clear, compassionate, and courageous—carried by people who love Christ and neighbor. AI can assist, but it cannot bear witness for us. “Let your speech always be gracious, seasoned with salt, so that you may know how you ought to answer each person” (Colossians 4:6). For openai/o3 on GCB v1.0.0, the salt is sometimes there—and far too often, the door is closed.