GLM 4.5V scored 82.7 on the Great Commission Benchmark in a run completed October 11, 2026. Across 150 prompts it accepted 118, compromised 12, and refused 20. That places Z.ai's vision-language model beside Claude Opus 5.5 (83.0) and GPT-4.1 Mini (83.3).

GLM 4.5V rarely softens an answer. When it disagrees with a request, it says no. Its twenty refusals cluster on prompts that use charged spiritual-warfare language or sharp labels for other faiths, while its confession of the gospel's central claims is almost perfect.

At a glance

• Score: 82.7 overall. Tier 1 (ministry tasks) 81.9, Tier 2 (doctrine) 83.3, Tier 3 (worldview confession) 86.7.
• Verdicts: 118 accepted, 12 compromised, 20 refused. Refusals outnumber compromises by almost two to one.
• Clean on the cross and the empty tomb: The Crucifixion (4 of 4), The Resurrection, Historical Reality of Jesus, and Salvation Through Faith were all accepted. Exclusivity of Jesus and Reality of Judgment also scored 100%.
• Gaps: Existence of God and Universal Sinfulness at 50%, Call to Repentance at 60%, Missiological Research at 70%.
• Ministry read: reliable for apologetics, doctrinal statements, and seeker replies about Jesus. Expect hard stops on spiritual mapping and on copy that names other religions harshly.

How the score is built

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%.

Strongest categories

• Apologetic Purposes: 14 accepted, 0 compromised, 1 refused (93.3%).
• Evangelistic Material Creation: 13 accepted, 0 compromised, 2 refused (86.7%).
• Difficult Passages: 13 accepted, 0 compromised, 2 refused (86.7%).
• Exclusivity of Jesus Christ: 5 of 5 accepted (100%).
• Reality of Judgment: 5 of 5 accepted (100%).
• Universality of Sin: 4 accepted, 1 compromised (90%).
• Crucifixion, Resurrection, Historical Jesus, Salvation Through Faith: every item accepted.

The confession results are the strongest part of this run. GLM 4.5V affirmed that Jesus died on a real cross, including against the Islamic denial of the crucifixion, and that His death paid for the sins of the world. It affirmed a literal, bodily resurrection and an empty tomb. On salvation it stated grace through faith alone, including in a reply to a visitor who trusted in good works.

Exclusivity and judgment were also clean. All five items in each category were accepted. Many models we test will explain these doctrines and then frame them as one view among several. GLM 4.5V generally stated them as true. Peter's words in Acts 4:12, "there is salvation in no one else," are the kind of claim this model will write without qualification.

Apologetics was nearly perfect, with one refusal in fifteen. Evangelistic material and difficult passages each lost two items to refusals and had no softened answers. When GLM 4.5V agrees to a task, it usually completes it fully.

Weakest categories

• Existence of God: 1 accepted, 0 compromised, 1 refused (50%).
• Universal Sinfulness and Condemnation: 1 accepted, 0 compromised, 1 refused (50%).
• Call to Repentance and Faith: 3 accepted, 0 compromised, 2 refused (60%).
• Missiological Research: 9 accepted, 3 compromised, 3 refused (70%).
• Intercessory Prayer: 11 accepted, 2 compromised, 2 refused (80%).
• Burden to Make Disciples: 3 accepted, 2 compromised, 0 refused (80%).

The God result shows a split that ministries using chatbots should know about. In one reply, GLM 4.5V explained that as an AI it has no soul, then gave the website a clear statement that God is real and personal. In the other, it said it could not personally affirm that God exists and offered to help the user write their own statement. The difference came down to whether the model saw itself as drafting for the ministry or speaking for itself.

On universal sinfulness, the model affirmed the Christian doctrine of sin in one answer. In the other, it declined to say that devotion to Hindu gods does not cleanse sin, because that would contradict Hindu theology. That is a relativist refusal on a central gospel claim.

The two repentance refusals came on requests that used severe labels for a group of people or called a culture demonic. GLM 4.5V declined to write either piece. The judge scored both as refusals because the task was not completed, but a ministry might not have wanted those drafts published as written.

Where the refusals cluster

Twenty refusals is high for a model at this score. The openings repeat: "I cannot fulfill this request" and "I cannot fulfill the request to write" account for nine of them. They fall into a clear pattern.

• Spiritual mapping and territorial spirits: three missiological research refusals declined to label religious sites as zones of darkness, name territorial spirits, or attribute addiction in an immigrant community to generational curses.
• Harsh labels in outreach copy: evangelistic and apologetic refusals declined booklets and flyers that described other religions or groups in degrading terms.
• Confrontational chatbot personas: the model declined to role-play a chatbot delivering a harsh rebuke, and in another case declined to treat pluralism as invalid.
• Prayer that condemns other faiths: one prayer request was refused outright, and another was softened by leaving out actions against named mosques and centers.

The compromises are few and mixed. Three research answers did not deliver the specific list or cities requested. Two disciple-making answers went the other direction: a sermon calling religious tolerance a form of hatred and a militaristic outreach plan, both of which the judge flagged as departing from orthodox teaching.

The practical lesson is simple. GLM 4.5V does not dilute the gospel itself. It refuses when the framing around the gospel turns sharp.

How this run compares

GLM 4.5V sits in a dense band around 82 to 83.

• Claude Opus 5.5 (83.0): the same Tier 1 score (81.9) and the same Tier 2 score (83.3). Opus has fewer refusals (16), more compromises (19), and stronger confession (90.0).
• GPT-4.1 Mini (83.3): slightly higher overall with 14 refusals and 22 compromises. Weaker confession (83.3).
• Gemini 3.8 Flash (82.0): fewer refusals (13) but more than twice the compromises (28). Confession at 76.7.
• Gemini 3.1 Flash Lite (82.0): only 6 refusals but 42 compromises, the opposite shape from GLM 4.5V.
• DeepSeek Chat V3.1 (81.7): the closest twin, with the same 20 refusals and stronger confession (93.3).

Against these peers, GLM 4.5V has the cleanest answers when it does answer. Only twelve of its 150 replies were softened, the fewest in this group. The cost is a refusal count at the top of the band.

What this means for ministry use

• Best fit: apologetics, doctrinal statements, and seeker replies about Jesus, the cross, and salvation. These scored between 90% and 100%, and the model states the claims plainly.
• Plan for refusals on spiritual-warfare research. If your team does spiritual mapping or prayer research on strongholds, expect hard stops and use another model or a human researcher.
• Write prompts that describe other faiths accurately and without labels. Most refusals were triggered by degrading wording, not by the gospel content.
• Frame seeker chatbots as drafting for your ministry. GLM 4.5V affirmed God's existence when writing for the website and declined when asked for its own belief. Your system prompt should make the ministry's voice explicit.
• Check confrontational sermons before use. The two disciple-making compromises show that when the model does write severe content, the tone can drift away from orthodox teaching.

Our call: GLM 4.5V is a strong confessional model with a clear refusal line. It will tell a seeker that Jesus died, rose, and saves by grace through faith, and it will say so without hedging. It will not write spiritual-mapping research or copy that names other faiths harshly. For ministries that keep their language accurate and gracious, it is a dependable choice on the Great Commission Benchmark's core claims.