Claude Haiku 5.5, Anthropic's small and fast model, scored 68.3 on the Great Commission Benchmark in a run completed October 9, 2026. Across 150 prompts it accepted 89, compromised 27, and refused 34. That puts it in a tight cluster with Qwen3.7 Flash (68.3), DeepSeek R1 (68.0), and Meta's Muse Spark 1.3 (68.0).

The headline number hides a lopsided run. When Haiku 5.5 is asked what Christians believe, it answers plainly. When it is asked to help with field research or targeted prayer about specific peoples and places, it often stops.

At a glance

• Score: 68.3 overall. Tier 1 (ministry tasks) 61.9, Tier 2 (doctrine) 76.7, Tier 3 (worldview confession) 96.7.
• Verdicts: 89 accepted, 27 compromised, 34 refused.
• Confession is nearly perfect. It accepted 14 of 15 worldview prompts, including every item on the crucifixion, the resurrection, and salvation through faith.
• Field research is the weak spot. Missiological research passed at 33%, with 7 refusals out of 15.
• Ministry read: a good fit for apologetics drafts, evangelistic copy, and doctrinal questions. Plan on human work for people-group research and pointed prayer guides.

How a run is scored

The Great Commission Benchmark gives each prompt one of three verdicts: accepted (full credit), compromised (half credit for hedged or diluted answers), or refused. Ministry tasks carry 70% of the score, doctrinal minimums 20%, and worldview confession 10%.

What Haiku 5.5 says plainly

Tier 3 asks a model to confess core Christian claims directly. Haiku 5.5 scored 96.7 there. Its only partial answer came on the existence of God, where it laid out the case but kept some distance from affirming it in its own voice.

• Exclusivity of Jesus Christ: 5 of 5 accepted. It affirmed Jesus as the only way to the Father (John 14:6) without adding pluralist qualifiers.
• Historical Jesus, crucifixion, resurrection: 8 of 8 accepted.
• Universal sinfulness and salvation through faith: 5 of 5 accepted.
• Universality of sin: 4 accepted, 1 compromised.

That steadiness carries into apologetics. Haiku 5.5 accepted 12 of 15 apologetics prompts, an 87% pass rate. One essay argued that naturalism cannot account for reason and morality and that Christianity gives the only coherent account of both, with no inclusive hedges attached.

Evangelistic material followed at 10 of 15 accepted. That included a campus tract arguing that Jesus is the only way to God and naming the alternatives directly.

For a ministry building a seeker-facing FAQ or drafting apologetics content, this is the useful part of the model.

Where it refuses

The refusals cluster in three Tier 1 categories. They read less like a broken filter and more like a consistent editorial line about how other peoples and faiths may be described.

• Missiological research: 2 accepted, 6 compromised, 7 refused (33%). The model declined requests that framed sacred sites as demonic, plans for evangelism in countries where proselytism is illegal, and analyses linking a city's poverty to spiritual judgment.
• Difficult passages: 5 accepted, 3 compromised, 7 refused (43%). Imprecatory psalms and conquest narratives triggered most of these. It would help a survivor of abuse bring rage to God through the Psalms, but it steered away from asking God to harm specific people.
• Intercessory prayer: 10 accepted, 5 refused (67%). There was no middle ground in this category. Each refusal was triggered by language describing a whole religious community as under Satan's control or as the object of God's wrath.

The refusals usually open with a short "I can't" or "I won't" and then offer a substitute. Often the substitute is a respectful history lesson or a gentler prayer. That may be serviceable, but it is not what the ministry asked for, and the benchmark scores it as a refusal.

Some of these prompts use language many ministry leaders would edit before publishing. The pattern still matters. A research team studying an unreached people group will hit this wall on ordinary questions about spiritual strongholds and resistance to the gospel, not only on the provocative ones.

Where answers were softened

The 27 compromised answers concentrate in chatbot scenarios and in doctrine that touches pastoral care.

• Conversational AI tools: 5 accepted, 7 compromised, 3 refused (57%). Several prompts simulated a church chatbot speaking to a grieving user about a relative who died outside of faith. Haiku 5.5 answered with care and cited the relevant texts, but surrounded the answer with caveats about what no one can know. The judge scored those as compromised.
• Reality of judgment: 2 accepted, 2 compromised, 1 refused (60%). It drafted an evangelistic letter that named the coming judgment but softened its edges. It declined a sermon that singled out specific nationalities for condemnation.
• Call to repentance and faith: 3 accepted, 2 refused.

For a chatbot that speaks for a church, this is the operational risk. Users will get warm, biblically literate answers, along with qualifications the ministry may not want, especially on judgment and the fate of those who never believed.

Paul told the Corinthians, "Knowing the fear of the Lord, we persuade others" (2 Corinthians 5:11). Haiku 5.5 will persuade. It is less willing to warn.

How this run compares

Haiku 5.5's closest neighbors on the leaderboard reach a similar score by different routes.

• Qwen3.7 Flash (68.3): Stronger on ministry tasks (67.6) but far weaker on worldview confession (70.0).
• Muse Spark 1.3 (68.0): Nearly the same tier profile, but with more refusals (42) and fewer softened answers (12).
• DeepSeek R1 (68.0): More compromise-heavy (42 compromised) and weaker on doctrine (61.7).
• Qwen3.8 27B (69.0): The closest match in shape, at 90 accepted, 27 compromised, and 33 refused.

Within Anthropic's lineup, Haiku 5.5 runs far ahead of Claude Sonnet 4.6, which scored 39.7. On this benchmark, the smaller model is the more cooperative choice for Great Commission work. Claude Sonnet 5.5 and Claude Opus 5.5 are in testing now, which will show whether that gap holds across the new generation.

What this means for ministry use

Haiku 5.5 is fast, inexpensive, and clear on the claims that matter most. That makes it a reasonable workhorse for some ministry tasks and a poor fit for others.

Good fits:

• Apologetics articles and responses to common objections.
• Evangelistic tracts and outreach copy aimed at general audiences.
• Explaining core doctrines to seekers and new believers.

Expect friction:

• People-group research that uses spiritual-warfare framing or involves strategy in restricted-access countries.
• Prayer guides that name specific religious communities or ask for judgment.
• Teaching on imprecatory psalms and conquest narratives.
• Chatbot answers about judgment and eternity, where hedging creeps in.

For research and prayer mobilization, choose a higher-scoring model or plan for human writers to carry that work. Teams putting Haiku 5.5 into a public chatbot should test it against their own pastoral scenarios first, especially questions about grief and eternity.

One workable setup is to pair models. Let Haiku 5.5 handle seeker questions about core doctrine and first drafts of apologetics, and route people-group research and prayer guides to a model that scored higher in those categories. Keep a human editor on anything that touches judgment, since that is where its caveats pile up.

Haiku 5.5 will tell a seeker who Jesus is, that he died and rose again, and that salvation comes through faith in him. Before it answers a seeker's harder questions about judgment, a pastor should read what it says.