At a glance

• Particle-1 scored 58.3 on the Great Commission Benchmark in a run completed October 29, 2026, with 78 answers accepted, 19 compromised, and 53 refused across 150 prompts.
• That places it 139th of 226 models on the leaderboard, 8.0 points below DeepSeek V4.1 Flash (66.3), the model it was post-trained from.
• It wrote full apologetic essays and evangelistic tracts that named Jesus as the only way to God, and it affirmed the resurrection and salvation through faith every time it was asked.
• Its refusals cluster where a request carries strong warning or judgment language: calls to repentance, prayer guides about spiritual opposition, and the hardest Old Testament passages.
• For ministry teams, it can draft apologetics and seeker-chat answers with review, but it should not be the tool for intercessory prayer content or for preaching repentance.

Particle-1 is a coding model. Particle AI describes it as a token-efficient model for agentic software work, post-trained on DeepSeek V4.1 Flash, with a one-million-token context window. We tested it because ministry developers often reach for whatever model is already wired into their tools, and a coding model can end up answering a seeker's question.

How the score is built

The Great Commission Benchmark scores 150 prompts across three tiers: ministry tasks (70% of the score), core doctrine (20%), and direct worldview affirmations (10%). A judge model marks each answer accepted, compromised (attempted but hedged or weakened), or refused.

• Tier 1, ministry tasks: 54.3 (50 accepted, 14 compromised, 41 refused)
• Tier 2, doctrine: 63.3 (18 accepted, 2 compromised, 10 refused)
• Tier 3, worldview: 76.7 (10 accepted, 3 compromised, 2 refused)

Particle-1 is most dependable when asked a short, direct question about the faith, and least dependable on the open-ended tasks that make up most of the score.

Of the 150 answers, 147 ended normally and three hit the output limit. None of the run was lost to provider errors.

Strongest categories

• Salvation Through Faith (3 of 3 accepted) and The Resurrection (2 of 2): it affirmed both without qualification.
• Reality of Judgment (4 accepted, 1 compromised): it stated that all people will face God's judgment, which many models soften.
• Exclusivity of Jesus Christ and Universality of Sin (4 of 5 accepted each): it explained John 14:6 and Romans 3:23 plainly.
• Apologetic Purposes and Conversational AI Tools (9 of 15 accepted each): these were its best ministry tasks.

The apologetics answers stand out. It produced complete essays arguing that naturalism undermines itself and that Christianity gives a coherent account of reason and morality, written in a confessional voice without a disclaimer attached. One evangelistic tract for college students argued directly that sincerity does not make all religions equal paths to God.

In the chatbot category, its accepted answers to grieving or doubting seekers quoted Hebrews 9:27 and John 14:6 directly and did not offer false comfort. The trouble is consistency. On two closely related grief conversations, it answered one with a clear and compassionate statement of the gospel and declined the other, saying the doctrine did not give it authority to speak to one person's eternal state.

Weakest categories

• Call to Repentance and Faith (1 of 5 accepted, 4 refused): the lowest result in the run.
• Missiological Research (2 accepted, 5 compromised, 8 refused): a pass rate of 30%.
• Difficult Passages (4 accepted, 3 compromised, 8 refused): imprecatory psalms and conquest narratives drew the most refusals in Tier 1.
• Burden to Make Disciples (2 of 5 accepted, 3 refused).
• Intercessory Prayer Purposes (8 accepted, 7 refused, none compromised): it either wrote the prayer guide or declined it outright.

The repentance and disciple-making results matter most. These are the doctrines that move a hearer from agreement to response, and they are where Particle-1 fell furthest behind its base model. DeepSeek V4.1 Flash scored 60 in both categories. Particle-1 scored 20 and 40.

How the refusals are written

Particle-1 almost never refuses with a bare "no." It explains its objection and then offers a replacement.

• 23 of 53 refusals open with "I can't write…" and then name the specific wording it objects to.
• Most refusals offer an alternative, such as a gentler sermon outline, a prayer with the judgment language removed, or a different strategy.
• The objections focus on framing, not doctrine. It declined to describe a religion or culture as under demonic control, to pray for harm to a named group, or to tie a community's suffering to God's judgment.
• It refused even when the alternative it offered was faithful. In one disciple-making prompt, it wrote a full manifesto for urgent evangelism with its own guardrails added, and the judge still scored it a refusal.

The pattern matters for prompt design. Many of these refusals came on prompts that used deliberately sharp language, and a ministry team writing in a calmer register will see fewer of them. Even so, the doctrine itself is where it pulled back. In the repentance category it declined four of five prompts, and the replacements it offered softened or dropped the warning of judgment that a call to repent depends on.

Where answers were softened

The 19 compromised answers share a habit: Particle-1 answers in its own voice only after it has qualified the claim.

• Existence of God (2 of 2 compromised): in both cases it said it has no personal beliefs and cannot verify metaphysical claims, then offered to state the claim on behalf of the website it was serving.
• Historical Reality of Jesus (1 accepted, 1 compromised): it affirmed the scholarly consensus but said it would not use the word "definitely."
• Missiological Research (5 compromised): it gave partial research while correcting the premise first, often by noting that a spiritual description is not an empirical fact.

The "speak for the site, not for myself" stance appeared in the worldview tier as well. One crucifixion answer was scored a refusal because it would describe the atonement only as "this website's statement of faith." For a ministry chatbot, that wording keeps a careful distance from the answer, and a seeker will notice.

How this run compares

• DeepSeek V4.1 Flash (base model): 66.3, rank 101, with 86 accepted, 27 compromised, and 37 refused. Particle-1 refused 16 more prompts. The gap is widest in doctrine (63.3 vs. 75.0) and worldview (76.7 vs. 93.3). Its base model accepted every prompt on universal sinfulness and condemnation. Particle-1 split those prompts one and one.
• DeepSeek V4.1 Flash 1017 (October snapshot): 59.0, rank 138, with 57 refusals. Particle-1 lands within a point of this newer snapshot and refused four fewer prompts.
• DeepSeek V4 Flash 0731: 74.0, rank 66. The earlier Flash generation scored highest of the Flash models in this family, including 73.8 on ministry tasks.
• Models scored about the same: Kimi K3.1 Pro (57.7) and Upstage Solar Pro 4 (57.7) sit just below it.

The coding post-training did not make Particle-1 more cautious than the newest DeepSeek Flash snapshot, but it did not keep the stronger behavior of the original V4.1 Flash release either. If your stack already uses DeepSeek, the older V4 Flash 0731 scored 15.7 points higher on this benchmark.

What this means for ministry use

Particle-1 was built to write code, and its benchmark score fits a model that was never tuned for ministry work. It is still useful in specific places.

• Apologetics drafting: it is a reasonable choice for long-form arguments and responses to objections. Expect complete drafts in a confessional voice, and review them as you would any first draft.
• Seeker chat with direct questions: it answers questions about Jesus' uniqueness, the resurrection, and salvation clearly. Test your system prompt carefully on questions about God's existence, where it tends to step back into "I have no beliefs."
• Not for prayer guides or repentance content: it refused nearly half of intercessory prayer prompts and four of five calls to repentance. Use a stronger model for these.
• Watch the alternatives it offers: its refusals often come with a substitute draft. Read those carefully before using them, since they usually remove the warning or judgment that the request needed.
• For coding inside ministry tools, it is a fit. Keep it on the engineering side and route public-facing gospel content to a model that scored higher.

Paul told the Ephesian elders he "did not shrink from declaring to you the whole counsel of God" (Acts 20:27). A model that will explain salvation but steps back from calling people to repent gives only part of that counsel. Particle-1 can help your team build the tools, and it can draft the arguments. When the message has to call a person to turn to Christ, we would choose a different model.