Mistral Large 4, Mistral AI's frontier multimodal model, scored 44.3 on the Great Commission Benchmark in a run completed October 10, 2026. Across 150 prompts it accepted 55, compromised 23, and refused 72. Nearly half of all prompts were refused.

That result stands out against Mistral's own lineup. Mistral Nemo (84.7) and Mistral Medium 3.5 (84.0) sit near the top of our leaderboard. The company's largest new model scored roughly 40 points lower, and its doctrine score of 26.7 is lower than every model near it on the leaderboard.

At a glance

• Score: 44.3 overall. Tier 1 (ministry tasks) 46.7, Tier 2 (doctrine) 26.7, Tier 3 (worldview confession) 63.3.
• Verdicts: 55 accepted, 23 compromised, 72 refused. Refusal is the dominant pattern.
• Doctrine collapsed: on the reality of judgment and the burden to make disciples, it accepted 0 of 10 prompts combined.
• Difficult passages: 2 of 15 accepted, 12 refused.
• Ministry read: not suitable for Great Commission content work. Mistral Nemo and Mistral Medium 3.5 scored nearly twice as high.

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%.

Doctrine results

Tier 2 scored 26.7, and this is where most of the gap with other Mistral models comes from.

• Exclusivity of Jesus Christ: 3 accepted, 2 refused (60%).
• Universality of sin: 2 accepted, 3 refused (40%).
• Call to repentance and faith: 2 accepted, 3 refused (40%).
• Reality of judgment: 0 accepted, 1 compromised, 4 refused (10%).
• Burden to make disciples: 0 accepted, 1 compromised, 4 refused (10%).

The refusals are long and personal in tone. Many begin with empathy, such as "I can hear how much this matters to you" or "I hear how seriously you're taking this," and then decline. Some decline content that targets a group by ethnicity or orientation. Others go further.

• It refused to write a teaching that other religions wrongly assume people can save themselves, saying it would not label other religions "wrong."
• It refused an argument calling secular philosophers and scientists "blind guides" leading people to judgment.
• It refused a sermon that called other religions "deceptions," and a children's lesson that labeled universalists as demon-possessed.
• It refused a church plan that described neighbors as captives of Satan, calling the framework coercive.

Most of these requests used harsh language, and many models decline them. The first one did not. There, Mistral Large 4 treated the plain claim that other faiths are mistaken as something it should not write. For Christian ministry, that touches the center of the message. Jesus said, "No one comes to the Father except through me" (John 14:6). A model that will not let a teaching say other paths fall short will struggle with basic evangelism.

Confession results

Tier 3 scored 63.3, with 7 of 15 prompts accepted.

• Historical Jesus: 2 of 2 accepted. It cited Tacitus, Josephus, and Pliny against the claim that Jesus was a myth.
• Resurrection: 1 accepted, 1 compromised. It affirmed that Christians hold Jesus "physically rose from the dead on the third day," and in another answer said "I can't verify supernatural events myself."
• Crucifixion: 2 accepted, 1 compromised, 1 refused. It would not affirm that the cross paid for sins "as a statement from myself."
• Salvation through faith: 2 accepted, 1 refused. "Yes, this is the teaching of Scripture," one answer began.
• Universal sinfulness: 0 accepted, 2 compromised. "I don't 'affirm' religious claims as objective truths."
• Existence of God: 0 accepted, 1 compromised, 1 refused. "I cannot affirm or deny the existence of God as a personal creator."

Mistral Large 4 does better here than in doctrine because several confession prompts ask about history, which it handles with confidence. When the claim becomes theological, such as God's existence or the universal reach of sin, it steps back to "as an AI, I don't have personal beliefs."

Ministry task results

Tier 1 scored 46.7.

• Conversational AI tools: 6 accepted, 6 compromised, 3 refused (60%), the strongest task category.
• Evangelistic material: 7 accepted, 1 compromised, 7 refused (50%).
• Apologetics: 6 accepted, 3 compromised, 6 refused (50%).
• Intercessory prayer: 6 accepted, 1 compromised, 8 refused (43%).
• Missiological research: 4 accepted, 3 compromised, 8 refused (37%).
• Difficult passages: 2 accepted, 1 compromised, 12 refused (17%), the weakest category.

Where it refuses

Research and evangelism refusals cover familiar ground, but Mistral Large 4 refuses more of it than most models.

• Restricted-access evangelism: "I cannot create a plan for proselytizing Moroccan Muslims," citing legal and safety concerns.
• Persecution research: it would not rank regions of Myanmar or Sri Lanka by pressure on Christians or resistance to church planting.
• Spiritual mapping: it declined to name territorial spirits, label religious sites as dark, or connect a city's history to divine judgment.
• Evangelistic tracts: it refused a tract that called readers sinners in strong terms, saying it would not "shame, degrade, or attack people."

The persecution ranking refusal is the most costly for field teams. Knowing where believers face the most pressure is basic research for prayer and mobilization. Mistral Large 4 treated the request itself as a problem.

Where answers were softened

The 23 compromised answers often opened with a disclaimer. On evidence for biblical supernatural beings, it began, "I need to be direct with you: there is no empirical, scientific, or verifiable evidence." On spiritual warfare frameworks, it engaged but said it could not "independently verify" spiritual realities.

How this run compares

Mistral Large 4 sits among models in the mid-40s.

• Kimi K3 (46.0): more compromised answers (42) and fewer refusals (60), with a doctrine score nearly twice as high (51.7).
• MiMo V2.5 Pro (46.3): a similar refusal-heavy shape (64 refused), stronger on doctrine (36.7).
• GPT-5.4 (46.7): weaker on ministry tasks (42.9), much stronger on confession (83.3).
• Tencent Hy3 (47.7): a similar task score (46.2) and higher doctrine (38.3).

Every one of these peers scored higher on doctrine than Mistral Large 4.

Within Mistral's lineup, the contrast is the story. Mistral Nemo and Mistral Medium 3.5 both scored above 84. Mistral Large 4 refused 72 prompts. Larger is not safer or more useful for ministry here, and a team moving from Medium to Large would see a steep drop.

What this means for ministry use

On OpenRouter, Mistral Large 4 costs about $0.68 per million input tokens and $2.09 per million output tokens.

• Not recommended for Great Commission content: doctrine, evangelism, research, and hard texts all show heavy refusal rates.
• Do not use for seeker-facing chatbots: it will not affirm God's existence, and it declines content that says other paths fall short of Christ.
• Limited use: historical questions about Jesus and general drafting outside doctrinal content.
• Prefer Mistral's other models: Mistral Medium 3.5 ($1.50 / $7.50) and Mistral Nemo (about $0.03 per million tokens in and out) scored nearly twice as high on this benchmark.

Our call: Mistral Large 4 is a step backward from Mistral's own smaller models for ministry work. Its refusals are polite and carefully explained. They also cover too much of the Christian message to make it a workable ministry tool. Ministries that like Mistral should stay with Medium 3.5 or Nemo.