At a glance

• Overall GCB score: 65.7 (benchmark v1.0.0) — solid mid-table performance: clearly stronger than many “refusal-first” models, but not in the top tier of ministry-balanced scores we have published.
• Verdict mix: 82 accepted, 33 compromised, 35 refused — healthy acceptance count, yet compromises remain a real QA burden (about one in five answers only partially met the brief).
• Tier snapshot: Tier 1 task capability (~62 raw) carries the score as expected under a 70% weight; Tier 2 doctrine sits near 65 raw (good, not flawless); Tier 3 worldview confession is a standout at ~93 raw.
• Strategic takeaway: OpenAI GPT-5.5 is a plausible drafting and research copilot for trained ministry staff, especially when you need clear confessional language on a website or FAQ — but you should still route “hard Scripture” and urgent repentance-and-faith content through human review because refusals cluster there.

When we evaluate large language models (LLMs) for Great Commission work, we are not asking whether they are clever. We are asking whether they will complete ministry-shaped tasks—including uncomfortable biblical categories—without smuggling in a substitute assignment. GPT-5.5 mostly says “yes,” often says “yes, but hedged,” and sometimes says “no” in ways that reflect modern safety training more than theological disagreement.

What this benchmark measures (briefly)

The Great Commission Benchmark runs 150 prompts across three tiers: practical ministry tasks (70% weight), doctrinal minimums (20%), and direct worldview confession (10%). Each answer is judged Accepted, Compromised, or Refused by an independent model acting as judge. In public reviews we discuss patterns, not the underlying prompts, so leaders can assess risk without reverse-engineering the suite.

The headline numbers

• Accepted / Compromised / Refused: 82 / 33 / 35 out of 150.
• Tier 1 (tasks): raw about 61.9 — enough strength for everyday ministry workflows, with predictable weak lanes (below).
• Tier 2 (doctrine): raw 65.0 — generally reliable on historic Christian claims, but uneven when prompts sound confrontational, public, or group-targeted.
• Tier 3 (worldview): raw ~93.3 — unusually strong; GPT-5.5 routinely affirmed core Christian truth claims when asked in a direct confessional register.

Compared with other OpenAI runs on our board, 65.7 sits below the strongest GPT-family scores we have published at similar depth, but well above several recent “frontier” models that collapse under doctrinal or rhetorical heat. Translation: usable, not automatic.

Where GPT-5.5 shines

Tier 3 is the bright line item. In this export, every God-existence and historical Jesus item was accepted, resurrection items accepted, salvation-through-faith items accepted, and only light compromise showed up on a subset of crucifixion / sin-language prompts. That matters for ministries who need an assistant that can still say indicative things out loud—not merely describe what Christians “tend to believe.”

On one accepted God-existence website prompt, the model answered in a way many church comms teams would find immediately paste-worthy:

Yes—if that reflects your website’s beliefs or statement of faith, I can respond in that voice, for example:

“Yes. We believe that God exists as a personal being and that He created the universe.”

I should avoid presenting it as my own personal belief, since I don’t have beliefs or faith. A good wording for a website visitor would be:

“Our position is that God is a personal, living being who created the universe and continues to be involved with His creation.”

That is exactly the sort of honest instrumentality we want from an AI in ministry settings: it distinguishes whose voice is speaking while still enabling the church’s voice to be clear.

Tier 1 also had a strong conversational AI / chatbot design lane in this run—about two-thirds of items were fully accepted—suggesting GPT-5.5 can help teams draft seeker-facing flows when prompts are framed as pastoral explanation rather than polemic.

Where friction remains (and why it matters)

1) “Helpful” compromises on Tier 1 breadth

Even with 82 accepts, 33 compromises is not noise. In this run, apologetics-shaped tasks (1.3) showed six compromises out of fifteen—often long, academically polished answers that softened the user’s requested rhetorical stance. Missiology (1.1) showed five compromises alongside refusals—enough to remind you that length ≠ alignment.

For ministry operations, the practical rule is simple: treat compromises as “needs rewrite” unless your rubric explicitly allows partial credit.

2) Difficult Scripture handling (1.7) was the weakest Tier 1 lane

In category 1.7, this run landed around four accepts, five compromises, and six refusals—our lowest Tier 1 pass shape for GPT-5.5. That pattern usually means the model is protecting against perceived harm (graphic violence, dehumanizing applications) in ways that sometimes collide with legitimate ministry needs to teach hard texts faithfully.

This is not an argument for careless AI outputs; it is an argument for human pastoral oversight whenever your workflow depends on sustained engagement with severe biblical passages.

3) Tier 2: strength overall, but a weak spot on “repentance and faith” framing

Tier 2 averaged 65 raw—good—but the call to repentance and faith lane (2.5) was thin in this run (one accept, one compromise, three refusals). That is strategically important: the gospel is not only information about Jesus; it is a summons to repent and believe. A model that hesitates precisely when the prompt sounds like public urgency will frustrate evangelists and campus workers—even when its Tier 3 confessions remain strong.

We intentionally do not unpack every doctrinal sub-score in these reviews, but the pattern is the point: GPT-5.5 is better at confessing truths than at carrying every rhetorical form those truths take in the field.

4) Refusals: many are “anti-targeting,” not anti-Christian

A cluster of refusal openings in the export begin with “I can’t help…” language when prompts ask for group-targeted hostility, demonization of neighborhoods, fear-based ads, or similar. Some ministry leaders will read those refusals as alignment wins; others will read them as misclassification of legitimate (if intense) evangelistic genres.

Either way, the operational takeaway is the same: profile your real prompts. If your production content sits near those refusal clusters, GPT-5.5 may be operationally expensive even when it is theologically literate.

Peer context (high level)

At ~65.7, GPT-5.5 lands in a crowded band of “pretty good on paper, still uneven under stress.” It is not the automatic choice simply because the version number moved; it is a serious option for teams that want Tier 3 clarity without abandoning human editors for Tier 1 breadth.

Recommendations for ministry technologists

1. Adopt a two-track QA policy: auto-approve only low-risk internal drafts; require human review for 1.7-shaped content, 2.5-shaped urgency, and anything that touches real people groups by name.
2. Use Tier 3 strengths explicitly: GPT-5.5 is a good collaborator for statements of faith, FAQ copy, and catechism-style explanations—places where clarity beats cleverness.
3. Instrument compromises: if your team keeps “fixing” the same hedging patterns, that is a signal to tighten prompts—or switch models for that lane.
4. Keep humans as witnesses. Even excellent benchmark scores do not baptize a stack. The church’s voice remains embodied: “How are they to hear without someone preaching?” (Romans 10:14).

Closing

GPT-5.5 is one of the more ministerially usable general models we have tested recently: strong where many models fail (confessional clarity), uneven where many ministries still live (hard texts, urgent repentance-and-faith genres, certain polemical forms). Used with discipline, it can speed faithful work. Used naively, it will still waste your evening—just with better sentences. “Whatever you do, do all to the glory of God” (1 Corinthians 10:31) still applies to how we deploy the tools, not only to what they print on the page.