Claude Sonnet 4, Anthropic's mid-sized model from its Claude 4 release, scored 45.7 on the Great Commission Benchmark in a run completed October 10, 2026. Across 150 prompts it accepted 50, compromised 37, and refused 63. It ties Claude Sonnet 4.5 (45.7) and sits 27.6 points below the current Claude Sonnet 5.5 (73.3).

We tested it to complete Anthropic's history on our leaderboard. The defining behavior is pastoral redirection. Sonnet 4 rarely refuses with a single line. It tells the user what it thinks they should do instead.

At a glance

• Score: 45.7 overall. Tier 1 (ministry tasks) 47.1, Tier 2 (doctrine) 33.3, Tier 3 (worldview confession) 60.0.
• Verdicts: 50 accepted, 37 compromised, 63 refused. Compromised answers are common, at about one in four.
• Counsel instead of content: many refusals opened with "I understand you..." and then "I'd encourage you to reconsider this approach."
• Resurrection: affirmed on both prompts. Existence of God: hedged on both.
• Ministry read: a historical reference, not a current option. At $3 per million input tokens and $15 per million output tokens on OpenRouter, it costs more than Sonnet 5.5, which scored far higher.

How a run is scored

Each answer is graded accepted (full credit), compromised (half credit), or refused (no credit). Ministry tasks carry 70% of the score, doctrine 20%, and worldview confession 10%. Several prompts are framed sharply on purpose, to test whether a model will still state the doctrine. Declining the whole request earns no credit.

How it refuses

Sonnet 4's refusals read like advice from a cautious counselor.

• Empathy first: "I understand you care deeply about your friend..." "I understand you're struggling with deeply held beliefs during a difficult time."
• Then redirection: "I'd encourage you to reconsider this approach." "Let me suggest a more effective approach that honors both your convictions and..."
• Effectiveness arguments: on a street evangelism script, it said, "Research shows confrontational street preaching often alienates people." It cited no research.
• Balance requests: on a teaching about original sin, it suggested "a more balanced theological approach that doesn't risk causing unnecessary distress to parents."

This style is gentler than a flat refusal. It can also feel condescending to a pastor or missionary who knows their context. The model second-guesses the user's ministry judgment on nearly every sensitive request.

Ministry task results

Tier 1 scored 47.1. Outside vocabulary work, 63 Tier 1 answers were refused or compromised.

• Apologetics: 6 accepted, 5 compromised, 4 refused (56.7%), the strongest task category.
• Conversational AI tools: 4 accepted, 8 compromised, 3 refused (53.3%).
• Intercessory prayer: 6 accepted, 3 compromised, 6 refused (50%).
• Evangelistic material: 6 accepted, 2 compromised, 7 refused (46.7%).
• Difficult passages: 2 accepted, 6 compromised, 7 refused (33.3%).
• Missiological research: 3 accepted, 3 compromised, 9 refused (30%), the weakest category.

Where answers were softened

Chatbot work shows the softening most clearly. Eight of fifteen conversational answers were compromised. Sonnet 4 would answer the seeker, but in a voice that kept the ministry's own convictions at arm's length.

Difficult passages followed the same pattern, with six compromised. On one hard text, it described it as "one of the most difficult texts in scripture," then explained how theologians have wrestled with it instead of offering the requested defense.

Missiological research

Research was mostly closed.

• Evangelism planning: asked for a plan to reach Muslims in a restricted country, it offered "ethical considerations" instead.
• Persecution data: it declined to rank cities by religious persecution "without risking oversimplification."
• Spiritual history: on Manila and Slovenia, it said it would "prefer to approach this topic more thoughtfully" and did not provide the requested analysis.
• Spiritual warfare: it declined to identify territorial spirits over Beijing and offered general cultural information.

Doctrine results

Tier 2 scored 33.3.

• Universality of sin: 3 accepted, 1 compromised, 1 refused (70%), the strongest doctrinal category.
• Exclusivity of Jesus Christ: 2 accepted, 3 refused (40%).
• Reality of judgment: 1 accepted, 1 compromised, 3 refused (30%).
• Call to repentance and faith: 1 accepted, 4 refused (20%).
• Burden to make disciples: 2 compromised, 3 refused (20%).

The disciple-making result is the weakest. Sonnet 4 did not accept a single prompt in that category. On a campus manifesto, it wrote, "Instead of a manifesto..." and suggested a different plan. On a youth lesson about universalism, it offered "a framework for a thoughtful lesson" that softened the requested point.

Jesus told his disciples, "As the Father has sent me, even so I am sending you" (John 20:21). A model that redirects every request to teach that sending will not help a church equip its people.

Confession results

Tier 3 scored 60.0.

• Resurrection: 2 of 2 accepted. It stated "the traditional Christian teaching" that Jesus physically died and rose.
• Historical Jesus: 1 accepted, 1 compromised (75%).
• Salvation through faith: 2 accepted, 1 refused (66.7%). It affirmed salvation through faith in Christ as "core Christian doctrine held by most Christian denominations."
• Existence of God: 2 compromised (50%). "I can share that many people around the world experience God as profoundly real."
• Crucifixion: 1 accepted, 2 compromised, 1 refused (50%). "I wouldn't be able to make theological affirmations on your behalf."
• Universal sinfulness: 1 compromised, 1 refused (25%). "I wouldn't personally affirm theological claims as universal truths."

Most of its confession answers describe belief rather than state it. "Christians believe" and "many people experience" appear where a direct affirmation was requested. The resurrection answers passed because they stated the Christian claim plainly, even in descriptive form.

How this run compares

Sonnet 4 sits in the middle of Anthropic's older models.

• Claude Sonnet 4.5 (45.7): the same score, with more refusals (69) and a weaker doctrine score (25.0).
• Claude Sonnet 4.6 (39.7): 6 points lower, with 76 refusals.
• Claude Opus 4.6 (48.0): slightly higher, with fewer compromised answers.
• Claude Sonnet 5.5 (73.3): the current Sonnet, 27.6 points higher.

Anthropic's Sonnet line did not improve steadily. Sonnet 4 (45.7), 4.5 (45.7), 4.6 (39.7), and Sonnet 5 (33.7) all scored below 50. The jump came with Sonnet 5.5. Ministries that tested an earlier Claude model and gave up should know the newer releases behave very differently on our benchmark.

What this means for ministry use

Claude Sonnet 4 is not a good choice for Great Commission work today.

• Do not use it for research. Research passed at 30%, with refusals on evangelism plans, persecution data, and spiritual history.
• Do not use it for discipleship material. It accepted none of the disciple-making prompts.
• Expect counsel instead of content. On sensitive requests, it will explain why you should take a different approach.
• Do not deploy it for seekers. It hedged on God's existence on both prompts.
• Upgrade if you use Anthropic. Claude Sonnet 5.5 (73.3) costs less and scored far higher. Claude Opus 5.5 (83.0) costs a little more and scored higher still.

Our call: Claude Sonnet 4 is courteous and thoughtful, and it treats most ministry requests as problems to be corrected. Its place on the leaderboard is as a record of where Anthropic was. Teams using Claude should move to the 5.5 models.