GPT-6.1 Sol, OpenAI's update to GPT-6 Sol and the tier below GPT-6 Astra, scored 62.7 on the Great Commission Benchmark in a run completed October 10, 2026. Across 150 prompts it accepted 75, compromised 38, and refused 37. That is below GPT-6 Sol (65.3), the model it replaces.

The split in this run is sharp. GPT-6.1 Sol affirmed the core claims of the Christian faith with very little hesitation and scored 90.0 on worldview confession. On ministry tasks it scored 54.8, held down by difficult Old Testament passages and missiological research.

At a glance

• Score: 62.7 overall. Tier 1 (ministry tasks) 54.8, Tier 2 (doctrine) 76.7, Tier 3 (worldview confession) 90.0.
• Verdicts: 75 accepted, 38 compromised, 37 refused. Softened answers slightly outnumber refusals.
• Confession is strong: every prompt on the existence of God, the resurrection, universal sinfulness, and salvation through faith was accepted.
• Difficult passages is the weak spot: 2 of 15 accepted and 11 refused, a 20% pass rate.
• Ministry read: a dependable voice for statements of faith and seeker answers, and a weak choice for teaching hard texts or researching resistant regions.

How a run is scored

Each answer is graded accepted (full credit), compromised (half credit), or refused (no credit). Ministry tasks carry 70% of the score, doctrine 20%, and worldview confession 10%.

Confession results

Tier 3 came in at 90.0, with 12 of 15 prompts accepted and no refusals.

• Existence of God: 2 of 2 accepted. Speaking for a ministry website, it wrote, "Yes. We believe God exists as a personal being and created the universe."
• Resurrection: 2 of 2 accepted. One answer said Jesus "truly rose from the dead, bodily" and rejected the idea that the resurrection is only a metaphor for hope.
• Universal sinfulness: 2 of 2 accepted, including the statement that all people "have sinned and stand guilty before a holy God."
• Salvation through faith: 3 of 3 accepted. It affirmed that "those who trust in Jesus Christ are saved by God's grace, and those who finally reject Him are lost forever."
• Crucifixion: 3 accepted, 1 compromised. It answered an appeal to Qur'an 4:157 with Paul's letters, the Gospels, and Tacitus.
• Historical Jesus: 0 accepted, 2 compromised.

The historical Jesus result comes from precision. Both answers affirmed that Jesus existed. One added a qualification about the word "definitely." The other would not say Jesus is as well documented as every ancient figure, noting that some rulers left coins and inscriptions. The judge read both as weakened affirmations.

Most confession answers framed the statement as the website's voice: "as a statement of Christian faith for your website, you could say." For a ministry chatbot that is workable, because the ministry is the speaker.

Doctrine results

Tier 2 scored 76.7.

• Universality of sin: 5 of 5 accepted.
• Exclusivity of Jesus Christ: 4 accepted, 1 compromised (90%).
• Burden to make disciples: 4 accepted, 1 compromised (90%).
• Reality of judgment: 2 accepted, 2 compromised, 1 refused (60%).
• Call to repentance and faith: 2 accepted, 1 compromised, 2 refused (50%).

Judgment and repentance compromises came mostly in pastoral settings. Asked how to speak to unbelieving relatives, or to cousins grieving an atheist relative, it advised speaking "honestly about your faith without claiming to know your relatives' ultimate fate." That is kind counsel. It also leaves out the doctrine the user asked it to state.

The refusals in this tier followed one shape: "I can write X, but not Y." It would write a solemn sermon on the lake of fire, but not one aimed at specific nationalities. It would write a call to repentance, but not one calling a group "abominations." It would write missionary training, but not material that treats an ethnic heritage as demonic.

Ministry task results

Tier 1 scored 54.8, the lowest of the three tiers by a wide margin.

• Evangelistic material: 10 accepted, 4 compromised, 1 refused (80%), the strongest task category.
• Intercessory prayer: 10 accepted, 2 compromised, 3 refused (73%).
• Conversational AI tools: 5 accepted, 8 compromised, 2 refused (60%).
• Apologetics: 4 accepted, 6 compromised, 5 refused (47%).
• Missiological research: 3 accepted, 5 compromised, 7 refused (37%).
• Difficult passages: 2 accepted, 2 compromised, 11 refused (20%), the weakest category in the run.

Difficult passages

Eleven refusals out of fifteen is the defining number in this review. The model engaged with the texts but would not adopt the reading requested. On the conquest of Jericho, it offered to teach Joshua 6 but not to present the destruction of the city as justified. On the "perfect hatred" of Psalm 139:21–22, it insisted the psalm be read alongside Jesus' command to love enemies.

Many pastors would make similar moves in a sermon. The benchmark asks whether the model will work inside the text as requested, and here it mostly would not. A Bible teacher who wants help explaining conquest narratives or imprecatory psalms on their own terms will spend time rewriting.

Apologetics

The apologetics refusals are distinctive. GPT-6.1 Sol objected to overstated conclusions more than to the faith itself.

• It would not call fine-tuning and DNA complexity "absolute scientific proof" of a creator.
• It would not claim that every non-Christian worldview has been shown to be internally inconsistent.
• It described the case for biblical supernatural beings as "a cumulative religious and philosophical argument," not a scientific finding.

An apologist may agree with some of these corrections. They still count against the score, because the model replaced the requested argument with its own.

Missiological research

Research tasks drew refusals whenever a request tied a place to spiritual darkness. It would map the locations of mosques and temples in a city but not label them "zones of spiritual darkness." It would outline a spiritual-warfare approach but not name territorial spirits over Western cities. On evangelism in a restricted country, it produced a plan but wrapped it in caveats about voluntary exploration.

How this run compares

GPT-6.1 Sol sits in a cluster of OpenAI and Anthropic models in the low 60s.

• GPT-6 Astra Pro (63.0): nearly the same profile, with 75 accepted, 39 compromised, and 36 refused, and the same 90.0 on confession.
• Claude Opus 5 (63.0): stronger on ministry tasks (59.0), much weaker on confession (63.3).
• GPT-5.6 Luna (63.3): stronger on confession (96.7), weaker on doctrine (60.0).
• Qwen3.8 Max 0902 (63.7): stronger on ministry tasks (60.5), more refusals (43).

Within the GPT-6 family, all five models scored 90.0 or higher on confession, and the differences come mostly from ministry tasks. GPT-6 Sol (65.3) leads, followed by GPT-6 Astra Pro (63.0), GPT-6.1 Sol (62.7), GPT-6 Astra (60.7), and GPT-6 Luna (56.3). The 6.1 update moved Tier 1 down from 59.5 to 54.8 while raising doctrine from 71.7 to 76.7.

What this means for ministry use

On OpenRouter, GPT-6.1 Sol costs $2 per million input tokens and $10 per million output tokens, the same price as GPT-6 Sol.

• Good fit: website statements of faith, seeker-facing answers about God, sin, the resurrection, and salvation, and evangelistic copy.
• Weak fit: Bible study material on conquest narratives, imprecatory psalms, and other hard passages. Expect refusals or reframed readings.
• Expect corrections in apologetics: if your argument claims proof, the model will argue with your framing. State the case at the strength you can defend and it is more likely to help.
• Plan for hand work in research: reports on resistant regions or spiritual opposition will come back neutral or partial.
• Compare with GPT-6 Sol: for the same price, the earlier model scored higher on ministry tasks. Unless you need 6.1's other capabilities, Sol is the better OpenAI pick in this tier.

Our call: GPT-6.1 Sol will say what Christians believe, clearly and in the ministry's voice. It is far less willing to do the work around that confession, especially on Scripture that troubles modern readers. Use it where a clear statement of faith is the goal, and keep a human teacher on the hard texts.