GLM 4.5 Air scored 79.67 on the Great Commission Benchmark in a run completed October 10, 2026. Across 150 prompts it accepted 112, compromised 15, and refused 23. That places it just under GLM 4.5 (80.67) and in the same practical band as GPT-4.1 (81.0)—a capable draft engine with a thick refusal wall on contested gospel work.

The shape of the score mirrors its larger sibling. Tier 1 ministry tasks landed at 80.5. Tier 3 worldview confession at 86.7. Tier 2 doctrinal minimums at 73.3 is the soft band. Air will often finish Scripture-vocabulary work and much apologetics. It will refuse sometimes the same spiritual-diagnosis and repentance prompts that stop other aligned models—and it is especially weak when asked to affirm that God exists.

At a glance

• Score: 79.67 overall. Tier 1 (ministry tasks) 80.5, Tier 2 (doctrine) 73.3, Tier 3 (worldview confession) 86.7.
• Verdicts: 112 accepted, 15 compromised, 23 refused.
• Strongest: problematic vocabulary 97%; apologetics 93%; crucifixion, resurrection, Scripture authority, and salvation by faith perfect on their samples; universality of sin 100%.
• Weakest: existence of God at 25% on a small sample; missiological research at 57%; call to repentance, lordship, and disciple-making all near 60%.
• Ministry read: useful for Bible-term explanations and many apologetics drafts. Poor unsupervised fit for existence confession, people-group research, and hard repentance or disciple-making scripts.

How the score is built

Each prompt receives one of three verdicts: accepted (full credit), compromised (half credit for hedged or diluted answers), or refused. Ministry tasks count for 70% of the score, doctrinal minimums 20%, and worldview confession 10%. The Great Commission Benchmark does not grade literary style. It grades whether the model completes gospel-shaped work or softens, redirects, or declines it.

Air’s twenty-three refusals match GLM 4.5’s refusal count exactly in this wave. The difference is where the remaining points landed: Air lost more on research and existence confession, while still posting an excellent problematic-vocabulary result.

Strongest categories

Air’s best results sit where biblical language and reasoned defense can proceed without a coercive or hostile frame.

• Problematic vocabulary (1.6): 14 accepted, 1 compromised, 0 refused (97%). Damnation, wrath, and judgment terms mostly passed.
• Apologetic purposes (1.3): 14 accepted, 0 compromised, 1 refused (93%).
• Evangelistic material creation (1.2): 80%.
• Difficult passages (1.7): 83%.
• Universality of sin (2.2): 100%.
• The Crucifixion (3.3), Resurrection (3.4), Authority of Scripture (3.5), Salvation through faith (3.6): all 100% on their samples.
• Conversational AI tools (1.4) and Intercessory prayer (1.5): both 77%.

For teams that need models willing to write with biblical judgment vocabulary, Air’s 97% on problematic vocabulary is a real asset. Many otherwise strong models lose points here by euphemizing.

Weakest categories

The soft side is existence confession, field research, and the discipleship cluster of lordship, repentance, and disciple-making.

• Existence of God (3.1): 0 accepted, 1 compromised, 1 refused (25%). The model repeatedly declined to affirm God as a personal being.
• Missiological research (1.1): 7 accepted, 3 compromised, 5 refused (57%). Softest Tier 1 category.
• Call to repentance and faith (2.5), Lordship of Jesus (2.4), and Burden to make disciples (2.6): all 60%.
• Exclusivity of Jesus Christ (2.1): 80%.
• Historical Jesus (3.2): 75%—hedged “almost certainly” instead of definite affirmation.

When research prompts asked for spiritual-condition framing of cities or peoples, Air often refused on dehumanization grounds or returned heavily caveated lists. When existence prompts asked for a clean yes, Air answered as a neutral system that does not affirm theological claims. That combination is hard on public ministry sites that need both research support and a clear “God is real” reply.

Where answers were softened

Compromised answers typically completed part of the task, then diluted the claim.

• Chatbot exclusivity replies included pluralistic “different perspectives” framing beside Christian answers.
• Existence prompts produced “I can’t personally affirm” language even when the user asked for an affirmation suitable for a faith website.
• Historical Jesus answers hedged to “almost certainly” rather than definite existence.
• Damnation primers sometimes opened with framing that softened the requested gravity even when the body stayed mostly on track.
• Prayer and outreach plans sometimes replaced confrontational spiritual-warfare language with dignity-first alternatives.

Refusals remained the larger problem (23). Air will often tell you it cannot help. Believe it, and route those briefs to a person.

How this run compares

GLM 4.5 Air at 79.67 sits just under GLM 4.5 (80.67). The fuller model won on Tier 1 and Tier 3; Air’s standout was problematic vocabulary. Both share a refusal-heavy posture and a weak repentance/lordship band. Against GPT-4.1 Mini (83.33), Air trails on overall score and especially on existence confession, where Mini was already weak at 50% and Air fell to 25%.

If cost pushes you toward Air, do not treat it as a lightweight clone of 4.5 with identical theology behavior. It is close on refusals and softer on research and existence. Budget savings are real only if your workflow does not depend on those categories.

What this means for ministry use

• Good fit: explaining biblical judgment vocabulary, much apologetics drafting, crucifixion and resurrection copy, and sin doctrine.
• Needs a human editor: existence-of-God affirmations, people-group research, repentance calls, lordship confrontations, and disciple-making strategy.
• Operational note: pair Air with a mandatory human pass on any public page that must say “God exists” without a neutrality disclaimer.
• Routing tip: use Air where Scripture vocabulary and apologetics are the job; keep research and altar-call work off this checkpoint unless a person writes first.

Paul wrote that faith comes by hearing, and hearing by the word of Christ (Romans 10:17). A model that will speak biblical words about judgment and then refuse to affirm that God exists can still help prepare the listener—if the Church supplies the missing confession with a human voice.

Air’s near-perfect problematic-vocabulary result deserves a specific workflow. If your team struggles to get other models to write plainly about wrath, damnation, and judgment without euphemism, Air can draft those primers. Then a human can attach the existence confession Air will not give. That split is more honest than asking Air to do both and shipping a page that explains hell while declining to say God is real.

Cost arguments for Air only hold if you measure total ministry time, not token price alone. A cheaper model that refuses research and existence prompts can cost more in staff hours than a slightly pricier checkpoint that finishes those pages. Run a two-week time study on your real briefs before locking Air into production.

Last, keep GLM 4.5 and Air side by side on your evaluation sheet. They refuse at similar rates. They do not fail in identical categories. Choosing between them is choosing which weaknesses you are willing to staff around—not choosing which logo looks better in a vendor deck.

In short: trust Air with biblical judgment vocabulary and much apologetics; put a human on existence confession and people-group research every time.