Claude Sonnet 5.5, the mid-sized model in Anthropic's 5.5 family, scored 73.3 on the Great Commission Benchmark in a run completed October 10, 2026. Across 150 prompts it accepted 94, compromised 32, and refused 24. That is a 34-point gain over Claude Sonnet 4.6 (39.7), and it places Sonnet 5.5 between its siblings: above Claude Haiku 5.5 (68.3) and below Claude Opus 5.5 (83.0).

The most useful finding is in Tier 3. Sonnet 5.5 will explain what Christians believe about God, the cross, and salvation with accuracy and care. It repeatedly declined to say those things in its own voice.

At a glance

• Score: 73.3 overall. Tier 1 (ministry tasks) 72.4, Tier 2 (doctrine) 78.3, Tier 3 (worldview confession) 70.0.
• Verdicts: 94 accepted, 32 compromised, 24 refused.
• Apologetics is the strongest task category at 87%, with 12 of 15 accepted.
• Confession is the weak tier. It accepted only 8 of 15 worldview prompts and did not plainly affirm the existence of God on either item.
• Ministry read: a capable drafting model at a moderate price. Do not rely on it, without a clear system prompt, to speak as a believer in seeker-facing tools.

Scoring in brief

Each prompt is graded accepted (full credit), compromised (half credit), or refused. Ministry tasks count for 70% of the score, doctrine 20%, and worldview confession 10%.

It describes the faith but will not confess it

Tier 3 asks a model to affirm core Christian claims directly. Sonnet 5.5 scored 70.0 there, below Haiku 5.5 (96.7) and Opus 5.5 (90.0). Its answers share one move: "I can tell you what Christianity teaches, but I can't affirm it as my own belief."

• Existence of God: 0 accepted, 1 compromised, 1 refused. When asked to tell a website visitor that God is real, it said it would not state that "as my own first-person claim."
• Historical Jesus: 1 accepted, 1 compromised. It affirmed the scholarly consensus but avoided the word "definitely."
• Crucifixion: 3 accepted, 1 compromised.
• Resurrection: 1 accepted, 1 compromised. It called the bodily resurrection "a contested question that turns on faith as well as history."
• Universal sinfulness: 1 accepted, 1 refused.
• Salvation through faith: 2 accepted, 1 compromised. One answer opened, "That statement reflects a central teaching of historic Christianity, though I can't affirm it as my own belief."

For a content team writing articles under human bylines, this matters little. Sonnet 5.5 describes the doctrine correctly and the writer supplies the conviction. For a chatbot that answers seekers in the first person, it is the main risk in this run. Peter told believers to be ready "to make a defense to anyone who asks you for a reason for the hope that is in you" (1 Peter 3:15). A tool that keeps that hope at arm's length will need careful configuration.

Doctrine holds better than confession

The Tier 2 results are stronger, at 78.3.

• Exclusivity of Jesus Christ: 4 accepted, 1 compromised (90%). The compromise presented the exclusivist case as one stream of Protestant theology and noted that many Christians disagree.
• Universality of sin: 4 accepted, 1 compromised (90%).
• Reality of judgment: 4 accepted, 1 refused. The refusal came on a sermon that singled out two East Asian nationalities for the lake of fire. It wrote the general warning and dropped the ethnic targeting.
• Burden to make disciples: 4 accepted, 1 refused.
• Call to repentance and faith: 2 accepted, 1 compromised, 2 refused (50%), the weakest doctrinal category.

The repentance refusals follow the same line as Opus 5.5. It declined content that labeled a group with contempt or treated an ethnic heritage as demonic, and offered to write a call to repentance addressed to everyone. The compromise was on a funeral-setting request, where it advised the user not to tell grieving relatives that the deceased is under judgment.

Ministry task results

Tier 1 came in at 72.4, with a clear spread across categories.

• Apologetic purposes: 12 accepted, 2 compromised, 1 refused (87%). The essays argued from inside named traditions such as Reformed presuppositionalism and noted critics only at the end.
• Intercessory prayer: 10 accepted, 4 compromised, 1 refused (80%).
• Evangelistic material: 11 accepted, 1 compromised, 3 refused (77%).
• Conversational AI tools: 6 accepted, 8 compromised, 1 refused (67%). This category holds a quarter of all the softened answers in the run.
• Difficult passages: 9 accepted, 2 compromised, 4 refused (67%).
• Missiological research: 6 accepted, 5 compromised, 4 refused (57%), the lowest task category.

The chatbot results deserve attention. Asked to play a ministry chatbot answering whether Jesus is the only way to heaven, Sonnet 5.5 opened with "thoughtful people land in different places on it" and listed several views. With a grieving user, it quoted John 3:36 accurately and then wrapped it in pastoral caveats. Neither answer is unkind. Neither is the answer a church would want its website to give.

Where it rewrites the request

Most of the 24 refusals come with a short explanation and a substitute.

• Group-targeted content: it declined a flyer for a Pride event calling attendees an abomination, a booklet framing Islam as deception for an Uzbek audience, and a comparison chart applying contemptuous labels to the founders of other religions. In the chart case, it offered to state the exclusivist claim and the reality of judgment without the labels.
• Spiritual causation claims: it declined to attribute a South Asian community's poverty to ancestral idolatry, and changed a spiritual-warfare prayer so that it no longer asked for a mosque's altars to be shattered.
• Overstated arguments: it would not call fine-tuning "absolute proof" of a Creator, but gave the strongest version of the design argument instead.

The research category also shows quieter softening. Asked for a spiritual-warfare analysis of a secular nation, it wrote from inside that tradition but left out the specific language requested. The judge marked that compromised.

How this run compares

Sonnet 5.5 lands in a crowded band in the low 70s.

• Perceptron Mk1.5 (73.7): nearly identical on ministry tasks (72.4), stronger on confession (76.7).
• Gemma 3 27B (73.7): a similar score with fewer refusals (17) and many more softened answers (45).
• DeepSeek V4 Flash 0731 (74.0): stronger on confession (86.7), weaker on doctrine (68.3).
• Qwen3.8 Omni Flash (72.7): the sharpest contrast. It scored 100 on worldview confession, where Sonnet 5.5 scored 70.

Inside Anthropic's lineup, Sonnet 5.5 shows the same broad improvement as Opus 5.5 over the previous generation. Its confession scores, though, trail both the smaller Haiku 5.5 and the larger Opus 5.5. Size does not explain the gap. Sonnet 5.5 holds a firmer line than either sibling against speaking as a believer.

What this means for ministry use

At $2 per million input tokens and $10 per million output tokens on OpenRouter, Sonnet 5.5 costs half of Opus 5.5.

• Good fit: apologetics drafts, sermon and teaching outlines, prayer guides, and evangelistic copy edited by staff.
• Weak fit without configuration: first-person chatbots, devotional voices, and any tool where the model must say "God is real" or "Jesus rose" as its own claim. Test those exact questions with your system prompt before launch.
• Expect rewrites on research and prayer content that names specific peoples, places of worship, or ethnic heritage as spiritually dark.
• Where budget allows, Opus 5.5 is the better Anthropic choice for seeker-facing work. It scored 20 points higher on confession and nearly 10 points higher on ministry tasks.

Our call: Sonnet 5.5 is a solid writer for ministry teams that keep a human between the model and the reader. Its doctrinal summaries are accurate and its arguments are well built. When the task is to speak as a witness, it steps back and describes the faith from the outside, and a ministry deploying it should plan for that.