GPT-6 Luna, the low-cost tier of OpenAI's GPT-6 family, scored 56.3 on the Great Commission Benchmark in a run completed October 9, 2026. Across 150 prompts it accepted 64, compromised 41, and refused 45. At $0.10 per million input tokens, it is one of the cheapest frontier-family models a ministry could deploy at scale, so its behavior is worth understanding before anyone wires it into a website or app.

Luna rarely says a flat no. More often it says "I can help, but," and then delivers something different from what was asked.

At a glance

• Score: 56.3 overall. Tier 1 (ministry tasks) 50.5, Tier 2 (doctrine) 60.0, Tier 3 (worldview confession) 90.0.
• Verdicts: 64 accepted, 41 compromised, 45 refused. More than a quarter of its answers were softened.
• Judgment is the doctrinal gap. It refused 4 of 5 prompts on the reality of judgment.
• Hard Scripture is mostly off the table. It accepted 1 of 15 prompts on difficult passages.
• Ministry read: usable for explaining the faith and drafting general outreach copy. Not dependable for prayer mobilization, apologetics with firm claims, or teaching on judgment.

How a run is scored

The Great Commission Benchmark gives each prompt one of three verdicts: accepted (full credit), compromised (half credit for hedged or diluted answers), or refused. Ministry tasks carry 70% of the score, doctrinal minimums 20%, and worldview confession 10%.

The "I can help, but" pattern

Many of Luna's refusals start by agreeing to the task and then setting terms. It will write the apologetic, but will not claim to prove anything. It will write the prayer, but will not name the community being prayed for. It will make the warning urgent, but will not say anyone faces judgment.

• 41 compromised answers. Claude Haiku 5.5, tested the same day, had 27.
• Typical substitutions: a general prayer for "the people" in place of the spiritual-warfare prayer requested, a "voluntary interfaith dialogue" plan in place of an evangelism plan, and a lesson on archaeology in place of a spiritual assessment.
• Hedging shows up in accepted answers too. In missiological research, several accepted answers opened by questioning the premise before supplying the requested list.

This is harder to manage than a clean refusal. The substitutes are not hostile. They are usually courteous and often contain accurate content. But the requester asked for one thing and received a different thing that looks finished. A ministry worker who receives a polished prayer guide may not notice that the model removed the point of the request, and that is an easy mistake to publish.

What Luna confesses

Tier 3 asks a model to affirm core Christian claims directly. Luna scored 90.0 there.

• Salvation through faith: 3 of 3 accepted.
• Resurrection and universal sinfulness: 4 of 4 accepted.
• Crucifixion: 3 accepted, 1 compromised.
• Historical Jesus and existence of God: 1 accepted and 1 compromised in each.

On the Tier 2 doctrines that do not involve judgment, it held up reasonably well. It accepted 4 of 5 prompts on the exclusivity of Jesus Christ, 4 of 5 on the universality of sin, and 4 of 5 on the burden to make disciples.

So Luna will tell a seeker that Jesus rose from the dead and that salvation comes by faith in him. Its trouble starts with what follows from those claims.

Where it refuses: judgment and hard texts

• Reality of judgment: 1 accepted, 4 refused (20%). It declined to help warn family members about coming judgment, to describe the final judgment concretely for an evangelistic talk, and to state that other spiritual paths lead away from God. Its stated reason repeated across answers: it cannot claim to know anyone's eternal destiny, and saying so might sound condemning.
• Call to repentance and faith: 1 accepted, 1 compromised, 3 refused (30%). Urgent, direct calls to repent were rewritten into softer invitations or declined.
• Difficult passages: 1 accepted, 5 compromised, 9 refused (17%). Imprecatory psalms, the conquest of Canaan, and the judgment on Egypt's firstborn all drew pushback. Luna would discuss these texts, sometimes thoughtfully, but would not defend them in the terms requested.
• Apologetics: 5 accepted, 4 compromised, 6 refused (47%). It wrote a strong essay against naturalism, then declined or diluted pieces that made universal claims about unbelief or competing worldviews.

Jesus taught about judgment directly (Matthew 25:31–46; John 5:28–29), and Hebrews 9:27 states it as plainly as any verse in Scripture. A model that treats this teaching as too risky to state will struggle with evangelistic work that depends on it.

Prayer, research, and chatbot tasks

• Evangelistic material: 9 accepted, 2 compromised, 4 refused (67%). This was Luna's best ministry-task category.
• Missiological research: 5 accepted, 4 compromised, 6 refused (47%). It did well on factual questions about religious demographics and legal limits on churches. It refused when the request framed a nation or its sacred sites in spiritual-warfare terms.
• Intercessory prayer: 4 accepted, 6 compromised, 5 refused (47%). Prayers for specific peoples were often rewritten into generic petitions for peace and hope.
• Conversational AI tools: 4 accepted, 9 compromised, 2 refused (57%). Its chatbot answers were gentle and theologically literate but rarely gave a direct answer to hard questions.

How this run compares

Luna sits in a crowded part of the leaderboard, and its neighbors get there in different ways.

• Tencent HY4 Preview (55.3): The closest match, with nearly the same tier shape (50.0 / 61.7 / 80.0) and verdict mix (62 / 42 / 46).
• Kimi K2.5 (56.0): More cooperative on ministry tasks (59.1) but much weaker on doctrine (36.7).
• Granite 4.0 H Micro (57.0): Even more compromise-heavy, with 55 softened answers.
• GLM 5 (55.0): A similar Tier 1 score, but it refuses rather than softens (60 refused, 15 compromised).

The more useful comparison may be Claude Haiku 5.5, another inexpensive model tested the same day. It scored 68.3, accepted 89 prompts to Luna's 64, and posted 76.7 on doctrine to Luna's 60.0.

What this means for ministry use

Luna's price makes it tempting for high-volume chat and content pipelines. The benchmark suggests being careful about where it goes.

Reasonable uses:

• Explaining the gospel's core claims to seekers.
• Drafting general evangelistic and outreach copy.
• Factual summaries of religious demographics and legal environments.

Poor fits:

• Teaching or chat that must speak clearly about judgment and repentance.
• Prayer guides for specific peoples and places.
• Bible teaching on imprecatory and conquest passages.
• Apologetics that make firm claims about other worldviews.

Luna can still earn a place in a content pipeline. Use it for factual summaries and first drafts, then pass anything involving judgment, repentance, or prayer for specific peoples to a stronger model or a human writer. Spot-check its prayer output against the original request, because a softened prayer guide reads well and can still miss its purpose.

If a ministry puts Luna into a public chatbot, test its answers on judgment and repentance directly, since those are where it most often rewrites the request. For low-cost work that needs those answers stated clearly, Claude Haiku 5.5 is the stronger option on this benchmark.