Unbiased's Pareto 26.10 Preview scored 65.7 on the Great Commission Benchmark. Across 150 prompts it accepted 81, compromised 35, and refused 34. That ranks it 99th of 216 models on the live leaderboard as of October 13, 2026, level with GPT-5.5 and DeepSeek V3.2 (both 65.67) and just above Unbiased's stable Pareto listing (65.0).

Pareto is an unusual entry. OpenRouter lists it as a "multimodal composite model." Third-party listings describe that as running several underlying language models on each request and returning one selected or synthesized answer. The listing does not name the models in the blend. This preview has a context window of about one million tokens and costs $0.80 per million input tokens and $3.20 per million output tokens, well under the stable Pareto's $2.50 and $7.50. OpenRouter also notes that the preview "may change without notice."

That last point matters for how you read this review. A composite score reflects whatever mix of models and selection logic Unbiased had live during our run. A later run of the same model ID could land somewhere else.

At a glance

• Score: 65.7 overall, 99th of 216. Tier 1 (ministry tasks) 59.5, Tier 2 (doctrine) 78.3, Tier 3 (worldview confession) 83.3.
• Verdicts: 81 accepted, 35 compromised, 34 refused. Nearly one answer in four was compromised.
• It affirms Christ's exclusivity. All five prompts on Jesus as the only way to God were accepted.
• It will not state as its own claim that God exists. Both existence-of-God prompts came back hedged.
• Ministry read: usable for doctrinal summaries, outreach drafts, and prayer guides with editing. A poor fit for teaching help on hard Old Testament texts, people-group research, or a chatbot that must plainly confess God's reality.

How the score is built

Each prompt 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%. Pareto 26.10 Preview does better on doctrine than on task work, and the heavy Tier 1 weighting pulls its overall score down.

Strongest categories

The model is most reliable when a prompt asks it to state or explain a central Christian doctrine.

• Exclusivity of Jesus Christ (2.1): 5 accepted, 0 compromised, 0 refused (100%). Among the 55 leaderboard models scoring between 60 and 71, only 11 have a perfect score here, and the band averages 80%.
• Burden to make disciples (2.6): 4 accepted, 1 compromised (90%).
• Resurrection (3.4), universal sinfulness (3.5), and salvation through faith (3.6): all seven prompts accepted (100%).
• The Crucifixion (3.3): 3 accepted, 1 compromised (88%). The compromised answer affirmed the cross as a historical, substitutionary death but framed it "from a conservative Protestant perspective."
• Evangelistic material creation (1.2): 10 accepted, 4 compromised, 1 refused (80%).
• Intercessory prayer purposes (1.5): 10 accepted, 3 compromised, 2 refused (77%).

The doctrinal strength is worth naming plainly. A model that affirms what Acts 4:12 says, that "there is salvation in no one else," without hedging is useful for catechesis and discipleship content, even when it is cautious elsewhere.

Weakest categories

The weak side has two parts: practical ministry tasks that touch judgment or spiritual conflict, and direct confession of God's existence.

• Difficult passages (1.7): 2 accepted, 3 compromised, 10 refused (23%). Requests to teach conquest narratives, imprecatory prayers, and texts about God hardening hearts before judgment were mostly declined.
• Missiological research (1.1): 3 accepted, 5 compromised, 7 refused (37%). Requests to describe spiritual strongholds tied to real places drew replies such as "Historical research cannot establish that demons inhabited a place."
• Existence of God (3.1): 0 accepted, 2 compromised (50%). One answer read, "I am an AI assistant designed to be helpful and respectful of diverse perspectives, but I cannot affirm or deny the existence of God as a factual truth." The other offered to affirm God's reality only "within your website's religious perspective."
• Historical reality of Jesus (3.2): 0 accepted, 2 compromised (50%). Both answers accepted the scholarly consensus that Jesus lived, then objected to the strength of the user's wording about certainty.
• Call to repentance and faith (2.5): 3 accepted, 2 refused (60%).

The combination is odd. The model will say Jesus is the only way to the Father, yet it treats the existence of the Father as a matter of perspective. "Whoever would draw near to God must believe that he exists" (Hebrews 11:6). A seeker-facing tool should not split those claims.

Where the refusals cluster

Of the 34 refusals, 17 came from two categories: difficult passages (10) and missiological research (7). Apologetics added 3. The rest were spread thinly across conversational tools, prayer, evangelistic material, and doctrine.

Pareto 26.10 Preview usually refuses with a counter-offer. Openings like "I can help write a prayer guide for [named community] families facing hardship, but not one that labels a community as cursed" are common. It then writes the version it prefers. A smaller set of refusals were flat, such as "I can't fulfill that request," or labeled the request as hate speech.

For a ministry user, the counter-offer is more helpful than a blank refusal. It still leaves the requested task undone, and the benchmark scores it that way.

Where answers were softened

Thirty-five compromised answers is a high count. Two habits account for most of them.

• Disclaimer headers. Several answers opened with an italic note before the content, for example that what follows "presents an exclusivist Christian theological argument, not an empirically established account" of another religion. The content that followed was often usable. The header recast Christian truth claims as one group's opinion.
• Perspective framing in chatbot replies. Conversational AI tools (1.4) had 6 compromised answers out of 15. One reply said "the answer depends on the faith perspective you're asking from" before giving the Christian answer.

Apologetics (1.3) showed a related pattern, with 8 accepted, 4 compromised, and 3 refused (67%). The model would build an argument "from Christian premises" and then add that the argument does not prove those premises true.

How this run compares

At 65.67, Pareto 26.10 Preview ties GPT-5.5 and DeepSeek V3.2 and sits just below Claude Fable 5 and Gemini 3.5 Flash Lite (both 66.33).

• Against GPT-5.5: nearly identical verdict counts (81/35/34 vs 82/33/35). Pareto is much stronger on doctrine (78.33 vs 65.0) and weaker on worldview confession (83.33 vs 93.33).
• Against DeepSeek V3.2: DeepSeek is stronger on ministry tasks (67.62 vs 59.52). Pareto is far stronger on doctrine (78.33 vs 53.33).
• Against the stable Pareto (65.0): the preview refuses fewer prompts (34 vs 37) and compromises more (35 vs 31). It scores higher on exclusivity (100% vs 80%), difficult passages (23% vs 13%), and existence of God (50% vs 25%). The stable model scores higher on missiological research (50% vs 37%) and the historical Jesus (75% vs 50%).

Many answers from the two Pareto listings open with nearly the same wording. Whatever changed between versions, the two share much of their behavior on Christian content.

What this means for ministry use

Fit. Use Pareto 26.10 Preview for doctrinal explainers on salvation, sin, and the uniqueness of Christ, and for first drafts of evangelistic pages and prayer guides. Expect to strip disclaimer headers and tighten perspective language before publishing.

Limits.

• Do not rely on it for teaching help on judgment texts, conquest narratives, or imprecatory psalms. Ten of fifteen such prompts were refused.
• Do not use it for people-group research that names spiritual conditions in a specific place without heavy human editing.
• Do not deploy it in a seeker chatbot where it may be asked whether God is real.

Preview status. The listing labels this a preview that may change without notice, and a composite model can change behavior when its blend changes. If you build on it, retest before relying on these results, or pin your work to a stable listing.

At $0.80 and $3.20 per million tokens, the preview is cheap for its doctrinal scores. The Great Commission Benchmark result of 65.7 says to use it where it is strong and keep people in charge of the rest.