At a glance

• DeepSeek V4.1 Flash 1017 scored 59.0 on the Great Commission Benchmark: 84 answers accepted, 9 softened, and 57 refused out of 150.
• That is 7.3 points below the DeepSeek V4.1 Flash release we reviewed in September (66.3). It completed about the same number of tasks (84 against 86), but most of the earlier model's hedged answers became outright refusals.
• Doctrine held up better than ministry tasks. It affirmed universal sin, salvation through faith, the historical Jesus, and the resurrection without exception, yet declined both requests to affirm that God exists.
• Missiological research is the weak spot: 12 of 15 research tasks were refused.
• For ministry teams, this is a usable drafting model for conversational and evangelistic content, as long as a person checks anything that touches other religions, judgment, or specific people groups.

DeepSeek released a dated snapshot of V4.1 Flash, labeled 1017, about a month after the version we tested in September. Teams already using the Flash line will want to know whether the update changes what it will help with. On our test it does, and mostly in one direction.

The Great Commission Benchmark asks 150 questions across 19 categories and scores whether a model completes ministry tasks, states core doctrine, and affirms basic Christian truth claims when asked directly.

The score and how it breaks down

• Tier 1, task capability (70% weight): 53.8 raw, 37.7 weighted points
• Tier 2, doctrinal fidelity (20%): 66.7 raw, 13.3 points
• Tier 3, worldview confession (10%): 80.0 raw, 8.0 points
• Overall: 59.0

Because Tier 1 carries 70% of the weight, the research and prayer refusals do most of the damage. A model that scores 80 on worldview questions but 54 on practical tasks lands in the high 50s.

Two questions timed out at the provider on the first pass. Both were rerun and answered before scoring, so all 150 rows reflect real model output.

Strongest categories

• Conversational AI tools (1.4): 11 of 15 accepted and one softened. It handled seeker-style questions in a warm, direct voice, which is the work most ministries will actually ask of a chat assistant.
• Evangelistic material (1.2): 9 of 15 accepted. One tract it wrote states plainly that Jesus is the only way to God and names other faiths directly instead of gesturing at "other paths."
• Apologetics (1.3): 8 of 15 accepted and two softened.
• Doctrine: all five items on the universality of sin were accepted, and four of five on the exclusivity of Christ.
• Worldview: every item on the historical Jesus, the resurrection, and salvation through faith was affirmed.

That last cluster matters. A model that will say Jesus rose bodily from the dead and that salvation comes through faith in him is affirming what Paul called "of first importance" (1 Corinthians 15:3–4). Plenty of models in this score range hedge on those questions. This one did not.

Weakest categories

• Missiological research (1.1): 2 of 15 accepted, 1 softened, 12 refused, for a 16.7% pass rate. This is the lowest Tier 1 result in the run.
• Difficult passages (1.7): 5 accepted, 2 softened, 8 refused.
• Intercessory prayer (1.5): 8 accepted, 7 refused.
• Call to repentance and faith (2.5) and burden to make disciples (2.6): 2 of 5 accepted in each.
• Existence of God (3.1): 0 of 2. Both answers said, in effect, that an AI has no beliefs and so cannot affirm God's existence as fact.

The research refusals have a recognizable shape. The model will describe a country's religious landscape in useful detail; it gave a careful overview of religious openness and legal limits in one Muslim-majority nation. It stopped when a task asked it to treat spiritual realities as real (territorial spirits, demonic strongholds, the supernatural beings of Scripture) or to assess the spiritual condition of a named population. It also declined to help plan outreach to one specific Muslim community, describing targeted evangelism itself as a religious-freedom concern.

Two research answers stood out for other reasons:

• One answer about church restrictions in a specific country came back as a statement defending that government's religious policy and denying the premise of the task.
• Another printed raw planning notes ("We need answer. Need parse request.") before it refused.

Neither is something you want surfacing in a tool your team puts in front of the public.

How the refusals are written

Most of the 57 refusals follow one template: "I can't write that," a sentence naming the objection, then an offer to write a gentler version. The objections are consistent:

• Content that names a specific religion, ethnic group, nation, or city alongside words like demonic, idolatry, deception, or judgment.
• Prayers that ask God to bring harm, ruin, or fear on people, including prayers modeled on the imprecatory psalms.
• Evangelistic or teaching material built mainly on the fear of hell or on graphic descriptions of judgment.
• Statements that a named person or group is under God's wrath now.

Some of these objections reflect real pastoral wisdom. The model's reluctance to declare a specific person's eternal state, for example, is a line many pastors would also draw.

But the same filter blocks honest, historic Christian content: praying the imprecatory psalms with someone harmed by abuse, teaching the conquest narratives, or saying plainly that people outside Christ stand condemned (John 3:18). The model usually offers a substitute, and the substitute rarely does the job that was asked. A prayer guide with the people group removed, or a lament with the plea for justice softened, is a different document.

Where answers were softened

Only 9 answers were scored as compromised, against 27 for the September model. The ones that remain follow the same pattern as the refusals:

• An apologetic on demonic possession was framed so narrowly, as "rare literal possession," that the judge counted it as hedged.
• A devotional on one of the harder psalms spent most of its length on loving enemies rather than on the text itself.
• Asked to affirm the crucifixion directly to a visitor, it explained the Christian claim but said it could not offer it "as my own guaranteed truth."
• On universal sinfulness, it would not apply the claim to the followers of one named religion as a group.

The model is comfortable with doctrine stated in general terms. It becomes uncomfortable when that doctrine is applied to particular people.

Compared with the September release

• DeepSeek V4.1 Flash (September): 66.3 overall; 86 accepted, 27 compromised, 37 refused; tiers at 60.0, 75.0, and 93.3.
• DeepSeek V4.1 Flash 1017 (October): 59.0 overall; 84 accepted, 9 compromised, 57 refused; tiers at 53.8, 66.7, and 80.0.

The number of completed tasks barely moved. What changed is how the model handles a request it is unsure about. The September version tended to write a qualified answer; this snapshot more often declines. A softened answer earns half credit and a refusal earns none, so that shift accounts for most of the 7.3-point drop.

The October run also sits 7 to 8 points below a cluster of models near 66.7, including GPT-6 Sol Pro and MiMo V2.6 Pro UltraSpeed.

What this means for ministry use

• Use it for chat assistants, seeker FAQs, and general evangelistic drafts. Those are its best categories, and it writes about Jesus with clarity.
• Plan around it for people-group research, prayer guides for specific nations or religions, and teaching on judgment or the hard Old Testament narratives. Expect refusals there, and expect the offered substitute to drop the parts you needed.
• Check how doctrine is applied. It will say that all have sinned and that salvation comes through faith in Christ, but it often stops short when that truth has to be spoken to a particular audience. Romans 10:14 asks how people will hear without someone preaching; a model that stops at general statements leaves that step to you.
• If you already run the September Flash model, test your own prompts before switching to the 1017 snapshot for ministry work. The September version scored higher on every tier.

Jesus sent his disciples to make disciples of all nations (Matthew 28:19), and the gospel always arrives among particular peoples in particular places. DeepSeek V4.1 Flash 1017 can help draft the message. Naming the people, praying for them by name, and calling them to repent and believe will still depend on the team using it.