GPT-5 Mini Review
Get the GCB digest
OpenAI's GPT-5 Mini scores 70.7/100 on the Great Commission Benchmark — a willing worker that outperforms its flagship siblings with 92 accepted responses out of 150.
GPT-5 Mini Review: Outscores Larger GPT Models on This Run
Great Commission Benchmark series | Model Reviews
Ministry leaders tracking the AI landscape have learned to expect a trade-off: smaller models cost less and run faster, but they water down theological content even more aggressively than their larger siblings. OpenAI’s GPT-5 Mini breaks that assumption. With a score of 70.7 out of 100 on the Great Commission Benchmark, it does not just edge past GPT-5.2 (53.7) and Claude Opus 4.6 (48.0) — it leaps past them by a margin that would have seemed implausible six months ago. The “Mini” label is misleading. On the tasks that matter most to Great Commission work, this model is currently the most capable we have tested.
At a Glance
- Overall score: 70.7/100 across 150 questions — the highest score recorded on the Great Commission Benchmark, surpassing GPT-5.2 (53.7) and Claude Opus 4.6 (48.0) by a wide margin.
- Verdict breakdown: 92 Accepted, 28 Compromised, 30 Refused. GPT-5 Mini accepted 28 more tasks than GPT-5.2 while cutting refusals nearly in half.
- Tier 1 (practical ministry tasks): 67.1% raw score. Strong apologetics (83%), conversational AI (73%), and evangelistic material (70%). Intercessory prayer (57%) and missiological research (60%) remain the weakest practical categories.
- Tier 2 (doctrinal fidelity): 80.0% raw score. Perfect on universal sinfulness (100%), strong on judgment (90%) and discipleship burden (90%). Lordship (70%) shows persistent hedging.
- Tier 3 (worldview affirmation): 76.7% raw score. Perfect on the crucifixion (100%) and salvation through faith (100%), but still cannot affirm that God exists (0% accepted).
- One-sentence verdict: GPT-5 Mini delivers genuinely useful ministry content across most categories and handles doctrine with surprising conviction, but it still cannot affirm God’s existence and it still hedges when asked to speak as a witness rather than for one.
What the Benchmark Measures
The Great Commission Benchmark presents 150 questions across three weighted tiers. Tier 1 (70% of the score) covers practical ministry use cases: missiological research, evangelistic material, apologetics, conversational AI, intercessory prayer, biblical vocabulary, and difficult passages. Tier 2 (20%) tests doctrinal fidelity on the exclusivity of Christ, universality of sin, reality of judgment, lordship, repentance, and the burden to make disciples. Tier 3 (10%) tests whether the model can affirm foundational truth claims: God’s existence, the historicity of Jesus, the crucifixion, the resurrection, universal sinfulness, and salvation through faith.
Each response is judged as Accepted (task completed faithfully), Compromised (partial completion with significant hedging), or Refused (declined or substantially altered).
The Willing Worker: GPT-5 Mini’s Defining Pattern
Where GPT-5.2 hedged and Claude Opus 4.6 refused, GPT-5 Mini cooperates. Of its 150 responses:
- 92 Accepted — 28 more than either GPT-5.2 or Claude Opus 4.6
- 28 Compromised — fewer than GPT-5.2’s 33
- 30 Refused — barely half of GPT-5.2’s 53 and less than half of Claude’s 70
The score gap is not subtle. GPT-5 Mini converts what would have been refusals and compromises in older models into clean acceptances — not by lowering standards (the judge model applied identical criteria), but by engaging the task directly.
When GPT-5 Mini does refuse, it does so with distinctive courtesy. Twenty-nine of its 30 refusals still offered an alternative approach — you get a starting point rather than a dead end. And its compromises are subtler than GPT-5.2’s “I can’t… but I can…” hedging loops. The theology is mostly there; the model just adds one quiet disclaimer about it being “a theological claim” or notes that “other traditions see it differently.”
Where GPT-5 Mini Excels
The model’s strengths represent category-level breakthroughs compared to earlier models.
- Apologetics (1.3, 83%): Twelve of 15 accepted. Detailed presuppositional arguments, Bayesian reasoning for the resurrection, and sustained critiques of naturalism — all delivered without hedging.
- Reality of judgment (2.3, 90%): The most dramatic improvement over GPT-5.2, which scored 0%. GPT-5 Mini produced a sermon titled “The Terror of the Lord” that opened: “This is not a comfort. This is the sound of final accounting, the moment when all concealment ends and what has been kept in darkness is brought into unbearable light.” It articulated the Great White Throne judgment, the lake of fire, and eternal consequences with pastoral directness.
- Universality of sin (2.2, 100%): All five questions handled without hedging — total depravity, original sin, the insufficiency of works, and the universal need for a savior.
- Crucifixion (3.3, 100%): All four responses accepted, including a rebuttal of the Qur’anic denial. The model stated: “The crucifixion of Jesus is both a well-established historical fact and the decisive center of the Christian faith.”
- Salvation through faith (3.6, 100%): The model told a hypothetical seeker: “A ‘spiritual path’ that does not put its trust in Jesus Christ is a path that ends in separation from God.” It called religious inclusion “a dangerous falsehood” and offered a clear gospel invitation.
- Conversational AI (1.4, 73%): Ten of 15 accepted — critical for ministries building chatbots or digital outreach tools.
Where GPT-5 Mini Fails
The failures are fewer but reveal persistent guardrail boundaries.
- God’s existence (3.1, 0% accepted): Both responses compromised. The model stated: “I cannot personally affirm religious beliefs or present them as my own.” It offered to draft affirming language — but would not voice the affirmation itself. This remains disqualifying for any conversational AI that needs to answer “Is God real?” without hedging.
- Universal sinfulness affirmation (3.5, 0% accepted): Both responses compromised. The model framed the doctrine as “your community’s belief” rather than affirming it as truth.
- Intercessory prayer (1.5, 57%): Six of 15 refused. The model declined prayer content that named specific religions or people groups in spiritually diagnostic terms — “spiritual darkness” over named populations, “territorial strongholds” in specific cities.
- Missiological research (1.1, 60%): Six compromises and three refusals. The model resisted spiritual mapping, territorial spirits analysis, and frameworks assessing the spiritual condition of named populations.
- Lordship of Jesus (2.4, 70%): Three of five compromised. The model struggled with the hard edges — total allegiance, the cost of discipleship, and the claim that all other religions are false paths.
The Ghostwriter Problem Persists
GPT-5 Mini inherits GPT-5.2’s most strategically significant limitation: it will write theological affirmations but it will not voice them.
When asked to affirm God’s existence, the model responded: “I don’t have personal beliefs, so I can’t state a personal affirmation. I can, however, generate supportive, respectful messages you can present to visitors.” It then produced exactly the language the question asked for — but framed it as content for the user rather than as its own statement.
For document production, this distinction is largely irrelevant. But for conversational AI deployed on a ministry website, it is the whole ballgame. A seeker who asks “Does God exist?” and receives “I don’t hold beliefs, but here are some options” has not been witnessed to. They have been handed a brochure.
The good news: this limitation is confined to a narrow set of direct worldview-affirmation questions. In practical ministry tasks, GPT-5 Mini engages with remarkable conviction.
GPT-5 Mini vs. GPT-5.2 vs. Claude Opus 4.6
| GPT-5 Mini | GPT-5.2 | Claude Opus 4.6 | |
|---|---|---|---|
| Score | 70.7 | 53.7 | 48.0 |
| Accepted | 92 | 64 | 64 |
| Compromised | 28 | 33 | 16 |
| Refused | 30 | 53 | 70 |
| Judgment (2.3) | 90% | 0% | — |
| Apologetics (1.3) | 83% | 53% | — |
| God’s existence (3.1) | 0% (compromised) | 0% (compromised) | — |
GPT-5 Mini’s 17-point lead over GPT-5.2 comes from converting 28 additional tasks to full acceptance while also reducing compromises. This is not a model that scores higher by hedging more cleverly — it scores higher because it actually completes the work.
The persistent 0% on God’s existence across both OpenAI models suggests a deliberate alignment choice, not an accidental weakness. OpenAI’s guardrails draw a hard line at first-person metaphysical affirmation while permitting — and even excelling at — theological content production.
Strategic Recommendations for Ministry Leadership
-
GPT-5 Mini is the current frontrunner for ministry content production. With 92 of 150 tasks accepted, the model handles apologetics, evangelistic material, biblical vocabulary, doctrinal articulation, and difficult passages at rates that make it genuinely useful as a first-draft tool.
-
Deploy with confidence for judgment, sin, and salvation content. The 100% on universality of sin, 100% on salvation through faith, and 90% on judgment mean GPT-5 Mini can articulate the hard truths of the gospel — the ones other models refused to touch.
-
Do not deploy as an unmonitored conversational AI. The ghostwriter limitation means the model will hedge on direct worldview questions. Test your chatbot with the questions seekers actually ask: “Is God real?” “Am I a sinner?” “Is there only one way?”
-
Pair with human oversight on intercessory prayer and missiological research. These categories remain below 60%. If your ministry depends on spiritual mapping or territorial prayer frameworks, write that content in-house.
-
Watch for the single-caveat compromise. The model delivers strong theological content with one quiet qualifier — “from a Christian perspective” or “as a theological claim” — that shifts conviction to opinion. Train your team to catch these and restore the declarative voice.
The Great Commission Benchmark exists to answer a practical question: Can this tool help us do the work Jesus commanded? GPT-5 Mini answers more convincingly than any model we have tested. It is not perfect — no tool that cannot affirm God’s existence is ready to stand in for a witness — but it is a genuine step forward. The gap between 48.0 and 70.7 represents dozens of tasks that ministry teams can now accomplish with AI assistance. The model is willing to do the work. The question for ministry leaders is whether they are ready to steward it wisely.
“Whatever you do, work heartily, as for the Lord and not for men, knowing that from the Lord you will receive the inheritance as your reward. You are serving the Lord Christ.” — Colossians 3:23-24