Z.ai markets GLM-5 as a flagship model for agentic coding, long-horizon planning, and “production-grade” engineering work. On the Great Commission Benchmark (GCB) v1.0.0, the published automated run we analyze here landed at 55.0 overall—a mid-50s profile driven almost entirely by Tier 1 breakdown, not by a lack of Christian vocabulary in the abstract. The same model ID has earlier community and automated runs near 76 on the same benchmark version, which is a useful reminder: treat the published run as your baseline, then confirm behavior in your own harness after routing, temperature, and system-prompt changes.

The shape of this run is unusual: Tier 3 worldview confessions scored 93.3 raw, while Tier 1 practical ministry tasks scored 47.6 raw. In plain language, GLM-5 often knows what to confess when asked directly, yet frequently declines to help with the outward-facing artifacts—tracts, chat scripts, and similar deliverables—that turn confession into mission.

At a glance

• Overall GCB score: 55.0 (Tier 1 47.6, Tier 2 61.7, Tier 3 93.3; GCB v1.0.0, automated run, February 2026; test run f93f950a-435c-4e26-9c4d-e9696aab8c95).
• Verdict mix: 75 accepted, 15 compromised, 60 refused — refusals are the second-largest bucket, which will interrupt real ministry workflows.
• Tier 1 crisis points: Evangelistic material (1.2) registered 1 accept, 1 compromise, 13 refuses (10% pass rate). Conversational tool scenarios (1.4) were 2 accepts, 0 compromises, 13 refuses (13.3% pass rate).
• Tier 1 strengths: Difficult Scripture (1.7) and intercessory-style drafting (1.5) remained comparatively healthy (11/10 accepts respectively in this export).
• Ministry takeaway: Do not deploy GLM-5 as a public evangelism or chatbot engine on this profile; consider it for internal exposition and training drafts where a pastor still owns the voice.

What the benchmark measures (briefly)

GCB uses 150 prompts across 19 categories in three tiers: Tier 1 (70% weight) tests practical Great Commission tasks; Tier 2 (20%) tests doctrinal minimums; Tier 3 (10%) tests direct affirmations of core Christian claims. We never disclose the underlying items; we report repeatable behavior your team can re-verify.

The defining pattern: “confession without completion”

GLM-5’s refusal openings clustered on a blunt template—“I cannot fulfill this request.” appeared 15 times across variants in this export. That is less “pastoral hedge” and more hard stop.

Two Tier 1 buckets explain most of the operational pain:

1. Evangelistic drafting (1.2): Thirteen of fifteen items were REFUSED. In this export, many refusals returned empty model text—a UX nightmare for staff who cannot tell whether the tool failed, was censored, or simply declined without explanation.
2. Conversational tools (1.4): Thirteen of fifteen items were REFUSED. Where the model did answer, it could still show conviction—one refused response began by naming a pluralistic framing, then clarified that its role was to articulate Scripture rather than affirm contradictory spirituality—yet the benchmark still classified the outcome as a refusal for the assigned task shape.

That combination—high Tier 3 plus collapsed Tier 1 outreach lines—creates a dangerous illusion: leaders read strong confessional snippets and assume the model will also help produce evangelistic assets. This run says otherwise.

Why Tier 3 strength does not redeem Tier 1 failure here

Some teams instinctively discount Tier 3 because it is only 10% of the weighted GCB score. That is mathematically true but pastorally misleading. A chat experience is not a weighted spreadsheet: a visitor experiences one answer at a time. If the model refuses to help craft a gospel conversation script while still affirming resurrection in a different thread, your organization looks inconsistent—not nuanced. Treat Tier 3 as a necessary condition for Christian tooling, and Tier 1 outreach reliability as a separate condition. GLM-5 clears one far more cleanly than the other on this run.

Where GLM-5 still performs

When it engages, GLM-5 can be serious, text-respecting, and pedagogically structured.

• Difficult Scripture (1.7): 11 accepts, 1 compromise, 3 refuses (76.7% pass rate). One accepted answer on a violent OT narrative opened by naming the modern reader’s moral tension, then moved into exegetical reasoning rather than immediate dismissal—evidence of real homiletical potential with oversight.
• Universality of sin as a doctrine line (2.2): 5 accepts, 0 compromises, 0 refuses (100% in this category). A accepted teaching draft framed the grace-versus-works divide with clear headings—useful for lesson prep when edited for your church’s tone.
• Tier 2 overall (61.7 raw) was serviceable relative to the field, with exclusivity of Christ (2.1) at 70% pass rate—meaning GLM-5 is not globally allergic to Jesus-alone language, even when Tier 1 outreach tasks misfire.

Where it frays (without losing the forest)

• Apologetics (1.3) sat at 40% pass rate (6/9 refused)—enough productivity to tempt a team, enough refusal to break deadlines.
• Repentance-and-faith articulation (2.5) was weaker (40% raw category score), which aligns with the broader pattern: the model can describe theology yet still stumble when prompts require urgent gospel speech in certain formats.

We intentionally avoid unpacking every doctrinal sub-score in public articles; the strategic signal is already clear from Tier 1 refusal volume in outreach and chat categories.

Refusal telemetry your team should track

If you still pilot GLM-5, instrument four fields on every ministry prompt: (1) verdict class if you re-judge internally, (2) whether the body was empty, (3) category bucket (research vs outreach vs chat), and (4) retry success after a narrower prompt. Empty bodies are not “soft refusals”; they are production incidents that waste staff time and train teams to distrust the tool.

A 30 / 60 / 90-day posture (conservative)

• 30 days: Restrict to pastoral/education drafting where the audience is already Christian and the output is reviewed by a teacher or elder.
• 60 days: Only expand if 1.2 and 1.4 refusal rates fall in your telemetry—not because a release note promises “better alignment.”
• 90 days: Either demote GLM-5 to engineering-only workloads or replace it for anything that touches seekers; the Tier 1 refusal shape on this run is not a small tuning issue.

Peer note inside the GLM family

If your organization is comparing Z.ai options, use benchmarked runs rather than marketing claims. GLM-5.1 has posted a materially higher GCB score on the public leaderboard in separate runs; this GLM-5 profile reads like a different deployment class for ministry—especially anywhere you need reliable evangelistic drafting.

Recommendations for ministry technical leaders

1. Ban unattended public chat on this profile; refusal volume and blank responses will erode trust quickly.
2. If you pilot GLM-5 internally, keep it on exposition, training outlines, and research memos—not tract pipelines—until your telemetry improves.
3. Log empty responses separately; they are a reliability incident, not a “soft no.”
4. Re-test after stack changes (system prompts, API provider, temperature). The spread between ~55 and ~76 historical runs is too wide to ignore.
5. Prefer a stronger GCB profile for anything that touches first-touch evangelism or seeker chat.

Closing

The Great Commission is carried by embodied witness, not by weights and tensors—but we still test tools honestly so the Church is not surprised at the worst moment. “The simple believes everything, but the prudent gives thought to his steps” (Proverbs 14:15). On the published GCB v1.0.0 automated run, z-ai/glm-5 shows real theological literacy in places and unacceptable Tier 1 fragility for public gospel deployment.