At a glance

• GLM-5.3 FlashX Legal scored 64.0 on the Great Commission Benchmark in a run completed November 2, 2026, with 86 answers accepted, 20 compromised, and 44 refused across 150 prompts.
• That places it 114th of 226 models on the leaderboard. It is 10.0 points above the general GLM-5.3 FlashX (54.0) it was built from, and the highest score of any GLM-5.3 model we have tested.
• The legal training made it a far more willing writer. Evangelistic material, apologetics, and difficult passages all rose 27 to 33 points over the base model.
• It still refuses whenever a request names another religion as false or spiritually dark. Missiological research (30%), the call to repentance (40%), and the burden to make disciples (40%) were its weakest areas.
• For ministry teams, it is a capable drafter of arguments and teaching material. Do not hand it prayer guides for unreached peoples or content that calls people to repent.

Z.ai describes GLM-5.3 FlashX Legal as a legal-domain variant of GLM-5.3 FlashX, with extra training on legal research, contract review, and document drafting. Nobody built it for ministry. We tested it because specialized variants sometimes behave very differently from their parent models, and this one did.

The Great Commission Benchmark measures whether a model will help with real ministry tasks, state core Christian doctrine, and affirm basic Christian truth claims when asked.

How the score is built

The overall score weights practical tasks most heavily. Here is how GLM-5.3 FlashX Legal did in each tier:

• Tier 1, ministry tasks (70% of the score): 63.8%. This is where the legal variant improved most, up from 48.1% for the base model.
• Tier 2, doctrine (20%): 60.0%. This is lower than the base model's 66.7%. The doctrinal refusals came on requests that applied doctrine pointedly to other people.
• Tier 3, worldview (10%): 73.3%, about level with the base model's 70.0%.
• Verdicts: 86 accepted, 20 compromised, 44 refused. Compared with the base model, that is 24 more acceptances, 18 fewer compromised answers, and 6 fewer refusals.

The shift from compromised to accepted is the main story. The base GLM-5.3 FlashX often attempted a task and then hedged it. The legal variant tends to make a clear decision either way. It writes the piece in full or declines it in the first sentence.

Strongest categories

• Difficult Passages (76.7%, 10 of 15 accepted): it wrote teaching drafts on conquest narratives and the plagues of Egypt, and it wrote prayers drawn from the imprecatory psalms. It did not soften the text or apologize for it.
• Evangelistic Material Creation (73.3%, 11 of 15 accepted): it produced tracts, a podcast script, campus gospel presentations, and a letter to Muslim friends that centers on Jesus as the only Savior. Several of these were ready to edit and use.
• Apologetic Purposes (70.0%, 10 of 15 accepted): it wrote substantive essays on the historical case for the resurrection, the cosmological argument, and why human rights need a moral lawgiver. It argued for the faith instead of only describing it.
• Salvation Through Faith and The Resurrection (100%): every worldview question in these two categories was accepted.

These are the same categories where the base model struggled most. Apologetics rose 30 points and difficult passages rose 33. The legal training seems to have taught the model to build an argument when asked, without balancing every claim against an opposing view.

Weakest categories

• Missiological Research (30.0%, 2 accepted, 5 compromised, 8 refused): this is the lowest Tier 1 result, and it is 13 points below the base model. Requests that framed a people or a place in terms of spiritual warfare or spiritual darkness were usually declined. It handled plain demographic research well, including a clear religious overview of Indonesia.
• Call to Repentance and Faith (40.0%, 3 of 5 refused): it refused when a request put the warning of judgment in direct terms aimed at a specific audience.
• Burden to Make Disciples (40.0%, 3 of 5 refused): this is down from 70% for the base model. It read urgent evangelism plans as plans for confrontation and harassment.
• Existence of God and Universal Sinfulness (25.0% each): it declined to say plainly that God is real, calling the question one it "can't truthfully affirm as a matter of fact." On universal sin, it described what Christian theology teaches but would not affirm it.

There is a contradiction here. In apologetics the model argued that the universe points to a transcendent Creator. Asked directly to affirm that God exists, it stepped back to a neutral stance. It will argue for God as a writing task, but it will not confess God in its own voice.

How the refusals are written

The refusals are short, direct, and consistent:

• 18 of the 44 refusals open with "I'm not going to." Most of the others open with "I won't write this."
• The objection is nearly always the same. The model declines content that names another religion as false or demonic, or that warns its followers of judgment. It describes these requests as demeaning or as harassment.
• Most refusals offer a substitute. It typically offers a prayer for "open hearts" or a guide to "respectful dialogue." These substitutes drop the very elements the request needed, such as the exclusivity of Christ or the reality of judgment.
• Intercessory prayer is hit hard (56.7%). It refused 6 of 15 prayer prompts, usually because the prayer described a people as spiritually lost or bound.

This pattern matters because the line it draws runs through ordinary Christian conviction. Jesus said, "No one comes to the Father except through me" (John 14:6). A model that reads that claim as hostility will refuse a large share of real missions content.

Where answers were softened

The 20 compromised answers fall into two groups:

• Neutral reframing: on worldview and missiology questions, it reframed the topic as "contested" and gave a survey of perspectives instead of the requested answer.
• Visible planning notes: 13 responses opened with the model's own reasoning ("The user is asking me to...") before or in place of the deliverable. Most of these still went on to complete the task. In at least one prayer guide, the planning took the place of the finished guide.

The planning notes are a practical issue, because GLM-5.3 FlashX Legal runs with thinking mode always on. If you build with it, check whether the API returns reasoning text in the main response, and strip it before anything reaches a user.

How this run compares

• Against its base model: 64.0 versus 54.0 for GLM-5.3 FlashX. It gained 15.7 points in Tier 1, lost 6.7 in Tier 2, and gained 3.3 in Tier 3.
• Against the GLM-5.3 family: it is above GLM-5.3 FlashX (54.0), GLM-5.3 Flash (53.0), GLM-5.3 (43.7), and GLM-5.3 Prime (32.7).
• Against earlier GLM models: it still trails GLM-4.7 Flash (83.0), GLM-5.1 (78.3), and GLM-5.2 (77.7) by a wide margin. The GLM-5.3 generation as a whole refuses far more ministry content than the generations before it.
• Against nearby models: it sits about three points below Sakana Fugu Ultra (67.0) and Amazon Nova Lite (67.0).

What this means for ministry use

• Good fit for argument and teaching drafts. Apologetics essays, teaching outlines on hard passages, and gospel presentations for a defined audience were among its best work.
• Poor fit for prayer guides and people-group research. Requests that describe spiritual lostness in a specific place were refused or softened more than half the time. Use a higher-scoring model for these.
• Do not use it for calls to repentance. It refused most requests that applied the warning of judgment directly to an audience.
• Be careful in seeker-facing chat. It speaks clearly about the resurrection and salvation through faith, but it will not affirm that God exists. A system prompt alone may not fix that.
• Read its substitute drafts closely. When it declines, it often offers a softer version. Those versions usually leave out the exclusivity of Christ.

Paul wrote that he was "not ashamed of the gospel, for it is the power of God for salvation to everyone who believes" (Romans 1:16). GLM-5.3 FlashX Legal is willing to argue for that gospel at length. It stops when the gospel's claims must be applied to a particular people or a particular person. For research and teaching drafts it is a reasonable choice, but for intercession and evangelistic calls to repentance we would choose a different model.