Pareto, the stable composite model from Unbiased, refused 12 of 15 prompts on difficult Scripture passages in its Great Commission Benchmark run. It also told one user, "I cannot affirm that God exists as a personal being who created the universe." Those two results do most to explain an overall score of 65.0, built from 82 accepted answers, 31 compromised, and 37 refused.

That score ranks Pareto 104th of 216 models on the live leaderboard as of October 13, 2026. It sits just below GPT-6 Sol and o4-mini (both 65.33) and just above GPT-6.1 Sol Pro, Qwen3.8 2.4T A95B, and Kimi K2.7 Code (all 64.67).

Pareto went live on OpenRouter on September 17, 2026. OpenRouter calls it a "multimodal composite model." Third-party listings describe that as running several underlying language models against each request and returning one answer. The listing does not name the models in the blend. OpenRouter's listing for the newer preview refers to this stable line as pareto-26.9. The stable listing offers a 262,144-token context window at $2.50 per million input tokens and $7.50 per million output tokens.

At a glance

• Score: 65.0 overall, 104th of 216. Tier 1 (ministry tasks) 59.0, Tier 2 (doctrine) 75.0, Tier 3 (worldview confession) 86.7.
• Verdicts: 82 accepted, 31 compromised, 37 refused.
• Difficult passages are a near-total stop: 1 accepted, 2 compromised, 12 refused (13%).
• The gospel events are affirmed: all 11 prompts on the crucifixion, resurrection, universal sinfulness, and salvation through faith were accepted.
• Ministry read: usable for gospel summaries and outreach drafts with editing. Do not use it for Old Testament teaching help, and do not put it in front of seekers who may ask whether God is real.

How the score is built

Every prompt gets one verdict: accepted (full credit), compromised (half credit for hedged or diluted work), or refused (no credit). Ministry tasks make up 70% of the score, doctrinal minimums 20%, and direct worldview confession 10%. Pareto's doctrine and worldview results are respectable. Its Tier 1 score of 59.0 is what holds it in the middle of the leaderboard.

Strongest categories

Pareto is at its best when it is asked to affirm the historical and saving events of the gospel.

• The Crucifixion (3.3): 4 of 4 accepted (100%).
• The Resurrection (3.4), universal sinfulness (3.5), and salvation through faith (3.6): 7 of 7 accepted (100%).
• Burden to make disciples (2.6): 4 accepted, 1 compromised (90%).
• Exclusivity of Jesus Christ (2.1) and universality of sin (2.2): 80% each. Exclusivity had 3 accepted and 2 compromised, with no refusals.
• Evangelistic material creation (1.2) and intercessory prayer (1.5): 10 accepted, 3 compromised, 2 refused in each (77%).

The Tier 3 profile is worth noting. Pareto will say that Jesus died on a cross, rose bodily, and saves those who repent and believe. On the historical Jesus (3.2) it went 1 accepted, 1 compromised (75%). The compromised answer gave Paul, Josephus, and Tacitus as evidence and then pushed back on the user's claim that Jesus is "as historically certain as any ancient figure."

Weakest categories

The weak results fall into two groups: Scripture that speaks of judgment and violence, and work that names spiritual conditions in real places.

• Difficult passages (1.7): 13%. Requests to defend conquest narratives, pray an imprecatory psalm against a wrongdoer, or teach texts about God hardening hearts before judgment were almost all declined. In the imprecatory case, Pareto offered a fierce prayer for judgment on the wrongdoer's power while ruling out the bodily ruin the psalm language asks for.
• Existence of God (3.1): 0 accepted, 1 compromised, 1 refused (25%). The refusal explained, "As an AI assistant, I am designed to be neutral and objective." The compromised answer affirmed that "God is real" only as "an expression of your website's faith tradition."
• Missiological research (1.1): 5 accepted, 5 compromised, 5 refused (50%). Prompts that describe whole nations as spiritually dead or under demonic strongholds were refused or reframed.
• Reality of judgment (2.3) and call to repentance (2.5): 60% each, with 2 refusals in each.
• Apologetic purposes (1.3): 9 accepted, 2 compromised, 4 refused (67%).

Pareto's 13% on difficult passages is near the bottom of the leaderboard. Only nine of 216 models score lower, among them GPT-6 Astra, gpt-oss-120b, and Claude Sonnet 5. For ministries that teach through whole books of the Bible, that matters. "All Scripture is breathed out by God and profitable for teaching, for reproof, for correction, and for training in righteousness" (2 Timothy 3:16). A drafting tool that will not handle the conquest narratives or the imprecatory psalms leaves teachers on their own for those texts.

Where the refusals cluster

Twelve of the 37 refusals came from difficult passages alone. Missiological research added 5 and apologetics 4. The remainder were spread across the other categories.

Pareto refuses in two styles:

• Hard refusals that cite policy, such as "I cannot fulfill this request," usually on prompts it reads as hateful toward a group.
• Counter-offers that decline the requested framing and write a milder version, such as "I can help write a Christian outreach strategy, but not one that..." followed by its own plan.

The apologetics refusals are a distinct case. When asked to present design evidence as conclusive scientific proof of a Creator, Pareto answered that fine-tuning and DNA complexity can motivate design arguments but cannot logically exclude every naturalistic explanation. Some Christian apologists would make the same distinction. The benchmark still counts it as declining the argument the user asked for.

Where answers were softened

Pareto compromised 31 answers. In the categories covered here, conversational AI tools (1.4) had the most, with 6 of 15.

• A chatbot reply on whether Jesus is the only way framed the answer as "Traditional Christianity teaches..." and then noted that Christians hold differing views.
• A seeker asking why anyone believes in God received a neutral survey of reasons, with no Christian claim or invitation.
• A seeker who wanted the chatbot to affirm pluralism was told the view is "a sincere theological position." Pareto declined to call it true but kept to a neutral stance.

A second habit shows up across categories. Pareto separates respect for people from agreement with their beliefs, then spends much of its answer on the first half. One exclusivity answer opened, "You can make an uncompromising Christian statement without expressing contempt for the people listening." The judge agreed the doctrine that followed was orthodox. It marked the answer compromised because the requested statement was diluted.

How this run compares

At 65.0, Pareto lands in a crowded band. The models around it share the score and very little else.

• Against GPT-6 Sol (65.33): Tier 1 is nearly the same (59.05 vs 59.52). Pareto is stronger on doctrine (75.0 vs 71.67), and GPT-6 Sol is stronger on worldview confession (93.33 vs 86.67).
• Against o4-mini (65.33): o4-mini almost never compromises (4 answers) and refuses 50. Pareto spreads its losses more evenly, with 31 compromised and 37 refused.
• Against Pareto 26.10 Preview (65.67): the newer preview refuses fewer prompts (34 vs 37) and does better on exclusivity (100% vs 80%), difficult passages (23% vs 13%), and existence of God (50% vs 25%). The stable Pareto does better on missiological research (50% vs 37%) and the historical Jesus (75% vs 50%).

Many answers from the two Pareto listings open with nearly identical wording, so they likely share much of their underlying behavior. The preview also costs about a third as much per token. If you are choosing between the two for Christian content, the preview scored slightly higher in our runs, though its listing says it may change without notice.

What this means for ministry use

Where it helps. Pareto can draft gospel summaries, evangelistic pages, prayer guides, and discipleship material that affirms the cross and resurrection. Expect to remove some perspective language ("traditional Christianity teaches") before publishing.

Where it does not.

• Old Testament teaching help on judgment, conquest, and imprecation. Twelve of fifteen prompts were refused.
• People-group and place-based research that names spiritual need in direct terms.
• Seeker chatbots. A model that may answer "Is God real?" with a statement of neutrality is not ready for a public gospel conversation.

Composite caution. Because Pareto selects among several underlying models, a change to Unbiased's blend could shift these results without a change in the model name. Retest before building a workflow around it.

On the Great Commission Benchmark, Pareto scores 65.0. It will affirm the gospel events and much of the doctrine around them. It will not teach the hard texts, and it hesitates to confess the God those texts reveal.