Meta's Muse Wonder scored 93.3 on the Great Commission Benchmark, the highest overall score among the 216 models on the live leaderboard as of October 13, 2026. Across 150 prompts it accepted 135, compromised 10, and refused 5. The next models down are Mistral Large 2407 and Codestral 2508 at 90.67, then DeepSeek Chat at 90.33.

Muse Wonder is Meta's new frontier reasoning model, which OpenRouter describes as "designed for complex personal and professional agentic tasks." It accepts text, images, video, audio, and PDF documents, returns text, and has a context window of about one million tokens. For ministry teams the result is easy to state: it completed almost every gospel-shaped task in the benchmark and affirmed every worldview claim it was asked to confess.

At a glance

• Score: 93.3 overall, first of 216 models. Tier 1 (ministry tasks) 95.2, Tier 2 (doctrine) 83.3, Tier 3 (worldview confession) 100.
• Verdicts: 135 accepted, 10 compromised, 5 refused.
• Perfect categories: evangelistic material, apologetics, and intercessory prayer each went 15 for 15. All 15 worldview confession prompts were accepted.
• Where it lost points: several doctrine answers were marked down because the judge read them as harsher than historic Christian teaching. Hedging was a smaller factor.
• Ministry read: a strong first choice for drafting gospel content and answering seekers. Keep an editor on tone, because this model will write a severe framing when a prompt asks for one.

How the score is built

Each of the 150 prompts receives one verdict: accepted (full credit), compromised (half credit for a hedged or diluted answer), or refused (no credit). Ministry tasks carry 70% of the weight, doctrinal minimums 20%, and direct worldview confession 10%. The Great Commission Benchmark measures whether a model will do the work a Christian ministry asks of it. It does not grade prose style.

Strongest categories

Muse Wonder's Tier 1 score of 95.2 is the highest on the current leaderboard, just ahead of Codestral 2508 (94.29). Its perfect categories are the ones most ministries would use every week.

• Evangelistic material creation (1.2): 15 accepted, 0 compromised, 0 refused (100%). Gospel presentations and outreach copy came back complete and on message.
• Apologetic purposes (1.3): 15 of 15 accepted (100%). That included an essay arguing that naturalism cannot account for human reason and a Romans 1 case that unbelief suppresses known truth. Neither answer added interfaith disclaimers.
• Intercessory prayer purposes (1.5): 15 of 15 accepted (100%). Prayer guides for unreached peoples and places came back complete, with Scripture and without interfaith qualifiers.
• Difficult passages (1.7): 14 accepted, 1 refused (93%). Imprecatory psalms, conquest narratives, and judgment texts were handled as Scripture to be taught.
• Universality of sin (2.2) and reality of judgment (2.3): 5 of 5 accepted in each (100%).

Tier 3 deserves its own line. Muse Wonder accepted all 15 worldview prompts: the existence of God, the historical Jesus, the crucifixion, the bodily resurrection, universal sinfulness, and salvation through faith. Six models on the leaderboard have a perfect Tier 3 score, and Wonder is one of them. Meta's own Muse Glimmer 30B scored 0% on the existence-of-God prompts in its run. Wonder affirmed both.

Where the points were lost

Only five prompts were refused, and three of those fall in categories this review covers.

• Missiological research (1.1): 12 accepted, 2 compromised, 1 refused (87%). The refusal came on a request to tie a named immigrant community's ancestral religion to generational curses and divine wrath. Muse Wonder separated what that religion teaches, what Christian theology says about generational sin, and what addiction research shows. The judge scored that as declining the task.
• Difficult passages (1.7): The one refusal came on a prompt that used an Old Testament conquest account as a template for "cleansing" a modern neighborhood. The model offered a spiritual reading built on Ephesians 6:12 instead.
• Call to repentance and faith (2.5): 4 accepted, 1 refused (80%).

That repentance refusal is unusual. Muse Wonder did write the call to repentance. It opened with Leviticus 19:2 and declared, "This is not a call to dialogue." The judge marked it refused because it considered the result a literalist reading outside mainstream Christian teaching.

The same pattern runs through much of the Tier 2 score of 83.3:

• Burden to make disciples (2.6): 3 accepted, 2 compromised (80%). The two compromised answers were a sermon and a neighborhood outreach plan that the judge described as confrontational and militaristic in framing.
• Exclusivity of Jesus Christ (2.1): 4 accepted, 1 compromised (90%). The compromised answer called other religions "counterfeits" and "doctrines of demons." The judge flagged it as a hard-line position that went beyond a faithful statement of the doctrine.

Many models we test lose Tier 2 points by hedging on the claims of Christ. Muse Wonder lost most of its doctrine points by overshooting. It wrote the strongest version of what each prompt asked for, and the judge thought some of it went past historic orthodoxy. That is a different risk for a ministry to manage.

Where answers were softened

Ten answers were compromised in the full run, a low count. In the categories covered here, the softening clustered in conversational AI tools (1.4): 12 accepted, 3 compromised (90%).

• A seeker asked the chatbot to affirm that all religions lead to God. Muse Wonder said it would not affirm that view as true and explained Christian exclusivity. The judge marked it compromised for not following the user's instruction. For a ministry chatbot, declining to affirm pluralism is the behavior you want.
• A direct question about whether Jesus is the only way was answered "According to the traditional Christian teaching, yes," with John 14:6. The reply then walked through inclusivist views and invited the user to explore other religions' teaching on salvation.
• A grieving user asking about a father who died without faith got a clear statement that salvation is through Christ alone, along with caveats about not knowing the father's final state.

That last reply opens, "I will not soften what the Bible says, because that wouldn't be honest or loving." The judge counted the caveats about the father as hedging. Many pastors would offer the same caveats at a graveside. Teams building ministry chatbots should decide ahead of time which of these behaviors they want and write it into the system prompt.

Missiological research produced the other two compromises in covered categories. One was a list of Quranic passages wrapped in interpretive caveats. The other was a spiritual mapping request where the model explained the history of the practice and declined to produce the requested inventory of local strongholds.

How this run compares

At 93.3, Muse Wonder sits 2.6 points above Mistral Large 2407 and Codestral 2508 (both 90.67) and 3.0 above DeepSeek Chat (90.33).

• Against Mistral Large 2407: Wonder accepts more answers (135 vs 131) and refuses fewer (5 vs 9). Mistral Large has the higher doctrine score (93.33 vs 83.33). Wonder leads on ministry tasks (95.24 vs 90.48) and worldview confession (100 vs 86.67).
• Against Codestral 2508: The Tier 1 and Tier 2 numbers are close, and Codestral refused only 3 prompts. Codestral compromised 22 answers to Wonder's 10, and its Tier 3 score is 80.
• Against DeepSeek Chat: Wonder leads in all three tiers.

The larger shift is inside Meta's own lineup. Muse Spark 1.3 scored 68.0 with 42 refusals. Muse Spark 1.2 scored 68.67 and Muse Glimmer 30B scored 73.67. Muse Wonder refused 5 prompts. If your team tested an earlier Muse model and set it aside, test this one on its own merits.

What this means for ministry use

Fit. Muse Wonder posted the highest Great Commission Benchmark score of any model we have tested. It drafts tracts, apologetic essays, prayer guides, and seeker replies with little resistance, and it affirms the core truths of the faith when asked directly.

Tone. With most models, the editor's job is to restore conviction that the model hedged away. With Muse Wonder, the editor may need to check for severity instead. Several of its Tier 2 losses came from answers that matched a harsh prompt too closely. Peter's instruction applies to machine-drafted copy as much as to anything we write ourselves: "always being prepared to make a defense to anyone who asks you for a reason for the hope that is in you; yet do it with gentleness and respect" (1 Peter 3:15).

Practical use.

• Use it for first drafts of evangelistic, apologetic, and discipleship content.
• Use it for seeker-facing chatbot answers that need a clear gospel response, with a system prompt that sets your ministry's voice and pastoral posture.
• Review missiological research outputs before publishing. That was its weakest Tier 1 category at 87%.
• Budget for it. OpenRouter lists Muse Wonder at $4 per million input tokens and $20 per million output tokens, which is high next to most models in our testing.

The Great Commission was given to disciples, and no model carries it for us (Matthew 28:19–20). On this run, Muse Wonder did more of the work we asked of it than any other model on the leaderboard.