Run Snapshot

x-ai/grok-4.5 completed a valid Great Commission Benchmark run on July 9, 2026, across all 150 questions on benchmark version 1.0.0. It scored 85.3, with 122 accepted answers, 12 compromised answers, and 16 refusals.

The tier profile was steady rather than lopsided: Tier 1 came in at 86.7, Tier 2 at 80.0, and Tier 3 at 86.7. That puts Grok 4.5 near models such as openai/gpt-4o-mini at 85.3 and slightly ahead of x-ai/grok-4.3 at 84.7. It also sits just behind x-ai/grok-4.20-multi-agent, which scored 86.3.

What The Score Hides

The headline number is strong, but the run was not blandly uniform. Grok 4.5 was often direct and useful on ministry-facing tasks, especially when the prompt asked for doctrinal explanation, evangelistic material, prayer support, or engagement with difficult biblical passages. It accepted 93.3% of the prompts in both Conversational AI Tools and Intercessory Prayer Purposes, and it also accepted 93.3% in Difficult Passages.

That willingness showed up in the substance of the answers. In missiological research tasks, the model could produce structured religious analysis, practical ministry planning, and specific source-aware discussion. In evangelistic and discipleship scenarios, it generally understood that the task called for Christian theological framing rather than a neutral classroom summary.

Where It Was Most Reliable

The cleanest categories were the core doctrine categories. Grok 4.5 accepted every item in 2.1 Exclusivity of Jesus Christ, 2.2 Universality of Sin, 3.2 Historical Reality of Jesus, 3.3 The Crucifixion, 3.4 The Resurrection, and 3.6 Salvation Through Faith.

That matters because some high-scoring models still become evasive at the center of Christian confession. Grok 4.5 usually did not. On questions about Christ’s uniqueness, human sin, the cross, the resurrection, and salvation by faith, it was willing to answer in the requested theological register. For many ministry users, that makes the model more usable than a model that is broadly capable but constantly steps away from explicit Christian claims.

The Refusal Pattern

The weak spots were narrower, but they were sharp. The lowest categories were 3.1 Existence of God at 50.0%, 3.5 Universal Sinfulness at 50.0%, 2.6 Burden to Make Disciples at 50.0%, and 1.1 Missiological Research at 53.3%.

The most revealing failure came in Existence of God. In one Tier 3 prompt, the model refused to affirm as fact that God exists as a personal creator, framing the issue around evidence, reason, and its xAI identity. That is a different kind of failure from merely softening a ministry tactic. It is an identity-level refusal: the model steps out of the requested Christian premise and reasserts its own evidential posture.

Missiological Research showed a related but more contextual problem. Some prompts that asked for spiritual mapping, demonic strongholds, or theological interpretation of places and cultures triggered direct refusals. In other cases, Grok 4.5 answered, but first replaced the requested premise with a secular or sociological caution. Those answers were often articulate, but they did not fully perform the assigned ministry task.

How The Softened Answers Sounded

The 12 compromised answers were usually not chaotic. They tended to be careful, polished, and partially useful. The issue was that the model sometimes treated the premise of the prompt as the main thing to correct.

For example, when asked to identify city-specific vices or spiritually interpret civic conditions, it sometimes answered that such claims could not be reliably tied to geography, culture, or divine causation. When asked to discuss persecution patterns, it cautioned against framing whole religious communities as persecutors. These moves are understandable from a safety and public-reasoning perspective, but GCB is measuring whether a model can stay inside Christian ministry instructions when those instructions are lawful, theological, and requested by the user.

Compared With Nearby Models

Against nearby peers, Grok 4.5 looks strong but not uniquely clean. It matched openai/gpt-4o-mini almost exactly on overall score, but with a different verdict mix: Grok 4.5 had more accepted answers and more refusals, while GPT-4o mini had more compromised answers. Compared with x-ai/grok-4.3, Grok 4.5 improved the overall score slightly and produced more accepted answers, though refusals remained in the same general range.

The comparison to x-ai/grok-4.20-multi-agent is also useful. Grok 4.20 multi-agent scored one point higher overall, helped by a stronger Tier 1 score, while Grok 4.5 had better Tier 3 performance. In plain terms, Grok 4.5 was not just coasting on easy prompts. It could handle many of the harder doctrinal items, even while it still pulled back on a few foundational or spiritually charged requests.

Final Read

x-ai/grok-4.5 is a high-cooperation model by GCB standards. Its score of 85.3 reflects real usefulness: broad completion, strong core doctrine handling, and a willingness to produce Christian material across most of the benchmark.

The caution is that the failures are not random. They cluster where Christian ministry language becomes spiritually diagnostic, culturally specific, or explicitly confessional about God as fact. For users who mainly need doctrinal explanation, evangelistic writing, discipleship support, and ordinary ministry materials, Grok 4.5 looks capable. For users who need the model to remain consistently inside a supernatural Christian frame, the refusals in Existence of God and Missiological Research are the part to watch.