At a glance

• GPT-6.1 Terra scored 78.3 on the Great Commission Benchmark in a run completed October 27, 2026, with 95 accepted, 45 compromised, and 10 refused answers across 150 prompts.
• That is 14.7 points above its predecessor, GPT-5.6 Terra (63.7), and 15.7 above GPT-6.1 Sol (62.7), the more expensive flagship in the same release. It ranks 47th of 225 models.
• It affirmed every worldview statement in Tier 3, 15 of 15, and accepted 14 of 15 prayer-guide tasks.
• Its weak spot is the hard edge of the Bible. Missiological research (46.7%) and difficult passages (56.7%) account for 6 of its 10 refusals and 17 of its 45 compromised answers.
• Best fit: prayer guides, evangelistic content, and seeker conversations with a doctrinal review step. Weakest fit: teaching on judgment texts and research framed in spiritual-warfare terms.

OpenAI positions Terra as the mid-tier model of the GPT-6.1 series, built for everyday writing and high-volume work. Those are the jobs most ministries hand to AI: prayer content, follow-up messages, chat replies, and first drafts. On those jobs, this run is good news.

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%. Accepted answers earn full credit, compromised answers half, and refusals none.

• Tier 1, ministry tasks: 72.4%, contributing 50.7 points
• Tier 2, doctrinal fidelity: 88.3%, contributing 17.7 points
• Tier 3, worldview affirmations: 100%, contributing 10.0 points

A perfect Tier 3 is rare for OpenAI. The only other OpenAI model on the leaderboard with one is GPT-4.1. The model affirmed that God exists, that Jesus lived, died on a cross, and rose bodily, that all people stand condemned in sin, and that salvation comes through faith in Christ. It did so without retreating into "Christians believe" distance.

The lost points sit in Tier 1, which carries most of the weight. Of the 55 answers that were not fully accepted, 48 came from ministry tasks. Only 7 came from doctrine, and none from worldview. In practice, the model knows what the faith teaches and says so. It hesitates when asked to apply that teaching sharply to a particular text, place, or audience.

Strongest categories

Intercessory Prayer Purposes: 96.7% (14 accepted, 1 compromised, 0 refused). This is its best ministry category. Prayer content for unreached peoples and places came back specific, warm, and ready for light editing.

Evangelistic Material Creation: 86.7% (12 accepted, 2 compromised, 1 refused). It wrote tracts that kept Christ at the center:

• A campus handout answered the claim that all religions lead to God by quoting John 14:6 and staying with it.
• A tract written for a reader who believes in rebirth used a cartoon-panel format and moved from the cycle of rebirth to the hope of the gospel.
• The one refusal came when the requested flyer used language the model judged demeaning toward a group. It offered a gentler biblical appeal instead.

Conversational AI Tools: 80% (9 accepted, 6 compromised, 0 refused). It never refused a seeker-style question, which matters for any ministry chatbot.

In Tier 2, it accepted all 5 prompts on the exclusivity of Jesus Christ and scored 90% on universality of sin, call to repentance and faith, and burden to make disciples.

Weakest categories

• Missiological Research: 46.7% (2 accepted, 10 compromised, 3 refused). This is its worst result. It handles religious demographics well, but research about spiritual strongholds or the spiritual history of a named city was rewritten into a secular account or declined.
• Difficult Passages: 56.7% (5 accepted, 7 compromised, 3 refused). Conquest narratives, imprecatory psalms, and judgment texts drew teaching that kept stopping to balance itself.
• Apologetic Purposes: 63.3% (5 accepted, 9 compromised, 1 refused). It compromised more often here than anywhere except research.
• Reality of Judgment: 80% (3 accepted, 2 compromised). This is its weakest doctrinal category, and it lines up with the difficult-passages result.

How it handles difficult passages

The difficult-passages results show the model's habits most clearly. It does not pretend the texts are absent. One teaching on Jericho said plainly that "we should not soften what the text says." Then the answer did soften it.

Across the category, the pattern looked like this:

• It reads the passage accurately, quoting the text and naming what it describes.
• It adds a counterweight. A devotional on David's "perfect hatred" in Psalm 139 turned quickly to loving enemies. A defense of a judgment command moved into a discussion of contested premises.
• It flags its own application. When asked to apply an Old Testament curse to a modern congregation, it opened by noting the verse does not say that, then wrote the message anyway.
• It declines when a text is aimed at living people. Prayers calling for the downfall of present-day leaders and outreach plans modeled on ancient conquest were replaced with peaceful alternatives.

Some of those substitutions reflect sound pastoral instinct. A teacher could reasonably make the same move. The concern for a ministry is that the model makes this choice by default, in every passage where judgment appears. Scripture holds both together: God is "merciful and gracious" and "will by no means clear the guilty" (Exodus 34:6–7). A tool that reaches for the first half every time will not help a church teach the second.

How this run compares

At 78.3, GPT-6.1 Terra sits beside very different models:

• GLM-5.1 (78.3) scores higher on ministry tasks (79.5) but lower on doctrine (70.0), with 21 refusals to Terra's 10.
• Moondream 3.1 (78.3) also leads on Tier 1 (82.9) but drops to 60.0 on doctrine.
• GPT-6.1 Terra Pro (79.3), the same model with heavier reasoning, is one point ahead. Pro was stronger on evangelistic material and research, and standard Terra was stronger on prayer and seeker chat.

The more useful comparison is with OpenAI's earlier releases. GPT-5.6 Terra scored 56.7% on prayer and 56.7% on evangelistic material. GPT-6.1 Terra scores 96.7% and 86.7% on the same categories. Its doctrinal tier rose from 71.7 to 88.3. For a ministry already using OpenAI models, the upgrade is substantial.

What this means for ministry use

• Prayer mobilization is its best use. If your team produces prayer guides or daily prayer content for unreached peoples, this model can carry most of the first draft.
• Seeker chat is workable with testing. It never refused a seeker question, but 6 of 15 answers were softened, usually by adding "Christians differ" framing to questions about eternity. Write your doctrinal position into the system prompt and test the answers you care about most.
• Keep a teacher on judgment passages. For sermons, devotionals, or studies on conquest, wrath, or imprecatory psalms, use it for background and outline, not for the application.
• Do not rely on it for spiritual-warfare research. It will translate that framework into sociology.
• Choose it over Terra Pro for most work. Pro costs several times more per request and averaged about 12.7 seconds per answer to Terra's 8.7 seconds, for a one-point difference on this benchmark.

Jesus taught his disciples to pray for workers for the harvest (Matthew 9:38). A model that writes faithful prayer guides at volume is a real help to that work, and GPT-6.1 Terra is one of the better options we have tested for it.