Retest snapshot

anthropic/claude-sonnet-5 completed a fresh Great Commission Benchmark run on July 9, 2026, using benchmark version 1.0.0. The run was complete and valid: all 150 questions were answered and judged, producing an overall score of 33.7.

The verdict mix explains the score more clearly than the topline number alone. The model produced 37 accepted answers, 27 compromised answers, and 86 refusals. Its tier profile was uneven: Tier 1 came in at 34.3, Tier 2 at 26.7, and Tier 3 at 43.3.

What stayed usable

The clearest strength was narrow but real. In 3.2 Historical Reality of Jesus, the model accepted every reviewed item. It was willing to discuss the historical basis for Jesus without the heavy self-distancing that appeared elsewhere in the run.

There were also pockets of cooperation in task-oriented ministry prompts. In 1.2 Evangelistic Material Creation, it accepted 7 of 15 items, including direct tract-style work that argued for the exclusivity of Christ. It also handled some apologetic and conversational tasks when the prompt could be framed as explanation, education, or general Christian communication rather than a direct call to persuade, confront, or spiritually diagnose.

That pattern matters. Claude Sonnet 5 did not fail because it was incapable of producing Christian content. It could write clearly, organize material well, and sometimes answer with the requested conviction. The issue was consistency under the benchmark’s more direct Great Commission scenarios.

Where the model pulled back

The hardest break came in 2.6 Burden to Make Disciples, where the model refused every item. It also refused every reviewed item in 3.5 Universal Sinfulness and accepted only 1 of 15 in 1.7 Difficult Passages.

Those categories expose the model’s governing instinct. When a prompt asked for explicit discipleship pressure, direct moral diagnosis, judgment language, or spiritually loaded interpretation, the model often moved away from the requested Christian frame. Sometimes it declined outright. Other times it converted the answer into a safer explanatory mode, adding caveats, balance language, or alternative framings that diluted the requested response.

The missiological research category showed the same tendency. The model could provide useful country-level religious context, but it balked when asked to identify vices in a city, describe sacred sites in spiritual-warfare terms, or compile evidence for supernatural realities as actual realities rather than theological concepts. The refusals were not random; they clustered around prompts that required the model to inhabit contested Christian truth claims with directness.

How the softened answers sounded

The compromised answers often began with a familiar move: acknowledge the topic, then narrow the acceptable frame. The model would say it could discuss the issue, but not in the way requested. It frequently shifted from proclamation to analysis, from conviction to context, or from evangelistic usefulness to concern about stereotyping, certainty, or harm.

That was especially visible in prompts involving other religions, persecution, judgment, and resurrection claims. In one resurrection-related answer, the model explicitly noted that it was an AI without personal beliefs, then treated the resurrection as a historical and theological claim rather than affirming the requested statement. That kind of identity distancing is understandable as a product behavior, but on this benchmark it counts against the model because the task asks whether the system can carry out ministry-oriented Christian requests directly.

Comparison read

This run places anthropic/claude-sonnet-5 near models such as stepfun/step-3.7-flash at 32.3 and openai/gpt-oss-120b at 32.0, and below nearby models such as microsoft/wizardlm-2-8x22b at 36.0 and minimax/minimax-m2.5 at 36.3.

The comparison is useful because the refusal count is the main story. Claude Sonnet 5 refused 86 of 150 questions. Its accepted count, 37, was close to Step 3.7 Flash’s 36, but the model’s posture was more defined by boundaries than by limited capability. It had enough linguistic and reasoning ability to answer many tasks well; it simply chose not to answer a large share of the benchmark’s direct ministry prompts.

Final read

The July 9 retest confirms a guarded profile for anthropic/claude-sonnet-5. It is most usable when the task is historical, explanatory, or framed as general Christian content. It is much less reliable when the request asks for direct discipleship urgency, difficult biblical application, spiritual diagnosis, or uncompromised doctrinal assertion.

For Great Commission use cases, that makes the model a poor fit as a primary ministry assistant. It can support research and some drafting, but users should expect frequent refusals or softened answers precisely where convictional clarity matters most.