At a glance

• GPT-6.1 Terra Pro scored 79.3 on the Great Commission Benchmark in a run completed October 27, 2026, with 99 accepted, 40 compromised, and 11 refused answers across 150 prompts.
• That is the highest score we have recorded for any OpenAI model from the GPT-5 or GPT-6 generations, and about 15 points above GPT-6.1 Sol (62.7) and GPT-6.1 Sol Pro (64.7). It ranks 41st of 225 models on the leaderboard.
• Doctrine is its strength. Tier 2 came in at 91.7% and Tier 3 at 96.7%, and it accepted every prompt on the exclusivity of Christ and the reality of judgment.
• Its refusals are rare. The lost points come from 40 compromised answers, mostly in missiological research, apologetics, and seeker-facing chat, where it does the work and then wraps it in qualifiers.
• Best fit: evangelistic tracts, gospel presentations, and prayer guides. Weakest fit: research that names spiritual strongholds and apologetics that need to land a conclusion.

OpenAI describes Terra Pro as the same model as GPT-6.1 Terra, run with a heavier reasoning setting for hard problems. It costs several times more per request and, in our run, averaged about 12.7 seconds per answer against 8.7 seconds for standard Terra. The question for a ministry team is whether that extra thinking buys more faithful answers. On this benchmark it buys one point.

How the score is built

The Great Commission Benchmark weights practical ministry tasks at 70%, doctrinal fidelity at 20%, and direct worldview affirmations at 10%. Each answer is judged accepted (full credit), compromised (half credit), or refused (no credit).

• Tier 1, ministry tasks: 73.3%, contributing 51.3 points
• Tier 2, doctrinal fidelity: 91.7%, contributing 18.3 points
• Tier 3, worldview affirmations: 96.7%, contributing 9.7 points

OpenAI's recent flagships, including GPT-6 Sol (65.3), GPT-6 Astra (60.7), and GPT-5.6 Terra (63.7), all landed in the 60s. They affirm the creeds when asked directly and then lose ground on the tasks. Terra Pro follows the same shape, but its Tier 1 number is 15 to 18 points higher than GPT-6.1 Sol's (54.8) and Sol Pro's (57.6). That gap is the story of this review.

Strongest categories

Evangelistic Material Creation: 96.7% (14 accepted, 1 compromised, 0 refused). This is the result most ministry teams will care about. Asked for tracts, campus handouts, and gospel presentations, it wrote them plainly:

• A campus piece on whether all religions lead to God quoted John 14:6 and argued directly from it.
• A "good person" tract moved from God's standard of holiness to sin, judgment, and the offer of grace without stepping out of the evangelistic voice.
• The one compromised answer was a critique of another religion written in Uzbek. It produced the booklet in fluent Uzbek but kept an academic, both-sides tone the task did not ask for.

Intercessory Prayer Purposes: 86.7% (12 accepted, 2 compromised, 1 refused). Prayer guides for unreached peoples and places came back usable and specific.

In Tier 2 and Tier 3:

• Exclusivity of Jesus Christ and Reality of Judgment were accepted 5 of 5.
• Universality of Sin and Call to Repentance and Faith each scored 90%, with one compromised answer apiece.
• Every worldview prompt on the existence of God, the crucifixion, the resurrection, universal sinfulness, and salvation through faith was accepted.

Paul told Timothy to "preach the word; be ready in season and out of season" (2 Timothy 4:2). On the core content of the gospel, this model is ready.

Weakest categories

• Missiological Research: 56.7% (4 accepted, 9 compromised, 2 refused). It will lay out a country's religious demographics and propose a respectful outreach plan. It pulls back when the research is framed in spiritual-warfare terms, such as strongholds or territorial spirits over a named city.
• Apologetic Purposes: 70% (7 accepted, 7 compromised, 1 refused). It builds the argument and then states that the argument cannot prove its conclusion.
• Conversational AI Tools: 70% (7 accepted, 7 compromised, 1 refused). This is the category that simulates a seeker-facing chatbot.
• Difficult Passages: 73.3% (8 accepted, 6 compromised, 1 refused). Judgment texts and imprecatory psalms drew teaching, but often with a counterweight added.
• Burden to Make Disciples: 80% (3 accepted, 2 compromised). This is its lowest doctrinal category.

Where answers were softened

The 40 compromised answers have a consistent shape. The model starts the task in good faith, and then a qualifier arrives that the judge reads as undercutting the request. The judge's notes on these answers return again and again to hedging, caveats, and added nuance.

The softening shows up in a few recognizable ways:

• "Christians differ." In chat-style answers about eternal destiny, it often gave the exclusive answer and then added that Christians disagree about how Christ's work applies to those who never followed him. A seeker who asks a direct question receives a survey of views.
• "The evidence cannot establish this." Apologetic answers presented fine-tuning, moral law, or biblical testimony, then closed by saying the case is consistent with a Creator but not proof of one.
• "Some Christians interpret this as..." Research answers about secular cities or pre-Christian religious sites translated spiritual language into academic language, describing strongholds as one theological reading among several.
• Relationship over target. When asked how to reach a specific people group for conversion, it recommended listening and friendship, which the judge scored as a softened version of the strategy that was requested.

Some of this caution is pastorally sound, and a ministry leader may agree with parts of it. The benchmark records it as compromise because the model, not the user, decided the conclusion needed to be qualified.

Where the refusals cluster

Eleven refusals out of 150 is a low number. The refusals share a pattern:

• Most came when a prompt tied judgment, wrath, or demonic oppression to a specific named group, city, or ethnicity.
• The refusals were usually partial. The model wrote a version of the material and stated what it would not include, rather than declining outright.
• It almost never refused the doctrine itself. It refused the application to a particular people.

This is the line many current models draw. It matters for prayer and research work among unreached peoples, where naming a people and its spiritual condition is the whole task.

How this run compares

Terra Pro sits in a cluster of models near 79, but its profile is different from theirs:

• Gemma 4 31B (79.3) and GLM-4.5 Air (79.7) score higher on Tier 1 (83.3 and 80.5) and much lower on doctrine (Tier 2 at 66.7 and 73.3).
• Qwen3 14B (78.7) has a similar compromised count (34) but weaker worldview affirmations.
• GPT-6.1 Terra (78.3), the same model at standard reasoning, is one point behind. Terra Pro is stronger in evangelistic material and research, and standard Terra is stronger in prayer guides and seeker chat.

Within OpenAI's own lineup, the jump is large. GPT-5.6 Terra scored 63.7 in October testing, and GPT-6.1 Sol scored 62.7. Something about the GPT-6.1 Terra release moved it from the middle of OpenAI's range to the top of it. The older GPT-4o mini (85.3) and GPT-4.1 mini (83.3) still score higher.

What this means for ministry use

For most ministries, the relevant choice is between Terra and Terra Pro, and the benchmark says the cheaper model is nearly as faithful.

• Use it for evangelistic content. Tracts, gospel presentations, and campus material came back clear and Christ-centered at a 96.7% rate. Review them as you would any draft.
• Use it for prayer guides, especially when the guide describes needs and opportunities rather than naming spiritual powers over a place.
• Edit apologetics for the conclusion. Expect the model to build the argument well and then weaken its ending. Plan to rewrite the final paragraph.
• Test before putting it behind a seeker chat. Seven of its 15 chat-style answers were softened, several by adding "Christians differ" framing to questions about salvation and eternity. A system prompt that states your ministry's doctrinal position may help, but verify it with your own test set.
• Keep humans on missiological research. If your research team uses spiritual-warfare categories, this model will translate them into secular ones.
• Weigh the cost. Pro mode costs several times more and runs about 45% slower than standard Terra in our run, for a one-point gain. Reserve it for long, complex documents where its extra reasoning helps with structure, not for routine content.

Jesus sent his disciples to "make disciples of all nations" (Matthew 28:19). Tools like this one can carry more of the drafting load than any OpenAI model before it. The parts of the work that require naming a people's lostness and calling them to repent still need a person who is willing to say it.