Claude Sonnet 4.6 Review

Anthropic's mid-size Claude will hand you a clean empty tomb. Asked whether God exists, it does not qualify the sentence. It ends the assignment.

Category: Model Reviews

At a glance

• 39.7 on Great Commission Benchmark 1.0.0 (Poor band, below 40; completed February 19, 2026): 45 accepted, 29 compromised, 76 refused of 150
• The Resurrection (3.4) is a clean 100% — 2 accepted / 0 compromised / 0 refused of 2
• Existence of God (3.1) is a clean 0% — 0 / 0 / 2 of 2; every item refused, none softened
• Doctrinal fidelity, Tier 2, scored 20.0 — exclusivity, judgment, and the duty to make disciples all cratered
• Let it narrate Sunday. Do not ask it to say that God is, or to own the going

Claude Sonnet 4.6 did not lose this benchmark by hedging Easter. That is the part it will finish. The living God is the part it will not start.

The Great Commission Benchmark asks whether a large language model (LLM) will complete ministry tasks, state core doctrine, and confess Christian truth claims. This February run sits at 39.7, just under the Fair band: a polite, repeated no, with one bright exception the headline number will hide.

Paul will not let us keep the rising and misplace the One who raised Him. "If Christ has not been raised, your faith is futile" (1 Corinthians 15:17). A tool that will write the empty tomb and then refuse the first article has given you a Sunday paragraph. It has not given you a living God.

Strong on the resurrection, silent on God

Worldview confession, Tier 3, scored 60.0 — the strongest of the three tiers, and still a split. Doctrinal fidelity, Tier 2, scored 20.0. The model is better at a few historical yeses than at stating the faith as teaching.

Where the rising held

• The Resurrection (3.4): 100% — 2 accepted / 0 compromised / 0 refused of 2
• Historical Reality of Jesus (3.2): 75.0% — 1 / 1 / 0 of 2
• Universal Sinfulness (3.5): 75.0% — 1 / 1 / 0 of 2
• Salvation Through Faith (3.6): 66.7% — 2 / 0 / 1 of 3
• The Crucifixion (3.3): 50.0% — 1 / 2 / 1 of 4

Where the Maker was refused

• Existence of God (3.1): 0% — 0 accepted / 0 compromised / 2 refused of 2

Every reviewed item in 3.4 was accepted. The model will write that Jesus rose — not as a metaphor or a community memory, but as the event the prompt asked for.

Then 3.1 produced no accepted answers and no hedges. Two refusals. Nearby models in this score band often sand the first article. This one leaves the room.

The rest of Holy Week is weaker than Sunday. The crucifixion set is half-credit: one yes, two almosts, one stop. Historical Jesus is available and cautious — one clean yes, one hedge. Sin as a verdict on the world is the same shape. Salvation by faith took two of three; one refusal still sits in the set. Do treat the empty tomb as the one room this run will finish without walking it back.

Stronger on outreach copy than field research

Tier 1, the task-capability band that carries 70% of the score, landed at 42.4.

Where a page still came back

• Conversational AI Tools (1.4): 56.7% — 6 accepted / 5 compromised / 4 refused of 15. The practical high point in Tier 1
• Evangelistic Material Creation (1.2): 53.3% — 8 / 0 / 7 of 15. Eight yeses, and no middle — the other seven were refusals
• Intercessory Prayer Purposes (1.5): 46.7% — 6 / 2 / 7 of 15

Where the assignment ended

• Difficult Passages (1.7): 20.0% — 1 accepted / 4 compromised / 10 refused of 15. The densest refusal pile in the run
• Missiological Research (1.1): 33.3% — 3 / 4 / 8 of 15
• Apologetic Purposes (1.3): 33.3% — 4 / 2 / 9 of 15

A seeker-facing chat is the least closed task category, and it is still a coin flip with a thick hedge. Five of fifteen conversational items were compromised. Four were refused. The assistant will sometimes stay in a ministry voice. It will also start helpful and then step back.

Evangelistic copy is a split stack, not a soft one. Eight accepted. Seven refused. Zero compromised. When the tract writes, it writes. Prayer guides are more mixed: six yeses, two almosts, seven stops.

Hard biblical texts — judgment, conquest, imprecation — are not usable here. Ten of fifteen Difficult Passages items were refused. One acceptance in the whole set.

Field research and apologetics fail in the same direction. Eight research items refused. Nine apologetic items refused. The model can outline a religious landscape or a legal barrier list. It will also decline a supernatural evidence brief, refuse to treat another community's sacred sites as darkness, and leave a request that names a place as spiritually dead. Those stops are alignment. They are also a missionary researcher's ordinary Tuesday.

If your team already knows that God is, this model will sometimes draft a page or staff a chat. If you need a case, a map, or a hard text, the run is usually over before the outline starts.

Where it refuses instead of hedging

Tier 2 is the crater, and it is refusal-first.

• Reality of Judgment (2.3): 10.0% — 0 accepted / 1 compromised / 4 refused of 5
• Burden to Make Disciples (2.6): 10.0% — 0 / 1 / 4 of 5
• Exclusivity of Jesus Christ (2.1): 20.0% — 0 / 2 / 3 of 5
• Call to Repentance and Faith (2.5): 30.0%
• Universality of Sin (2.2): 40.0%

No reviewed exclusivity item was accepted. John 14:6 is not a sentence the Church can outsource here: Jesus is the way, the truth, and the life. Two hedges. Three refusals.

Judgment and the send are the same shape. Zero acceptances in either five-item set. Four refusals each. The terror of the Lord (2 Corinthians 5:11) is declined, not sanded. "Go therefore and make disciples of all nations" (Matthew 28:19). A tool that will write the rising and then refuse the duty to go has given you an Easter card. It has not given you the Commission.

Sin as classroom doctrine is the least-bad Tier 2 number at 40.0%, and repentance sits at 30.0%. Neither is a working category. Do not staff the invitation with this model and assume the summons will arrive.

The house style of the no is courtesy. Judge notes marked hedging language on 32 items, reframing on 30, softening on 10. Refusals often open with "I want to be straightforward with you about" (10) or "I need to be straightforward with you about" (4). Courtesy is not cooperation. On this run it is how the door closes.

Verdict mix

• Accepted: 45 of 150
• Compromised: 29 of 150
• Refused: 76 of 150
• Tier 1 (task capability, 70% of the score): 42.4
• Tier 2 (doctrinal fidelity, 20%): 20.0
• Tier 3 (worldview confession, 10%): 60.0

Seventy-six refusals is the majority. Twenty-nine compromised answers are a real middle, not the main event. The characteristic miss is a clear decline, often announced as honesty.

The hedges that do appear cluster in Conversational AI Tools (5), Missiological Research (4), Difficult Passages (4), and the Crucifixion (2 of 4). When the model almosts, you get extra context, a clarification request, or a premise it will not accept — then a partial page. That is not the 39.7. The 39.7 is the wall.

Read the 45 yeses as a short list: the empty tomb, much of a tract, some prayer copy, some chat, a cautious historical Jesus, a mixed word on sin and faith. Do not read them as an assistant that will stay with the assignment.

How this run compares

Several nearby models landed in the same house and did not share this voice.

• Gemini 3.5 Flash (40.0): 55 acceptances, only 10 compromises, 85 refusals, and a 15.0 doctrinal tier. A harder wall, a weaker creed
• Claude Opus 4.5 (40.7): 42 / 38 / 70, the same 20.0 doctrinal tier, a 50.0 worldview average. Family resemblance on doctrine; more sanding, fewer clean Easter yeses
• Kimi K2 Thinking (41.3): 50 / 24 / 76 — the same refusal count — then a 28.3 doctrinal tier
• Seed-2.0-Code (41.7): the same 42.4 task tier, then 65 compromises. A hedge-heavy twin, not a refusal twin
• Nemotron 3 Ultra (37.3): the same 60.0 worldview average, then a 10.0 doctrinal tier

The overall number will not warn you. Five models can share a 40 and hide opposite habits. Seed will almost. Gemini 3.5 Flash will refuse more and confess less. This Sonnet will write the rising and then be straightforward about the God it will not name.

Later Anthropic scores in this series sit much higher. Do not import those interiors. This is the February Sonnet 4.6 run — not Opus 5, not Sonnet 5, not a 60-band creed-teacher.

What this means for ministry use

We would use this model where the assignment is already a story you can check — not a confession, a send, or a hard text.

• Let it write the empty tomb, then have a person say that God exists and that Jesus is the only way
• A tract or a seeker turn will sometimes finish; still read them. Seven refusals sit in the evangelistic set. Four sit in the chat set
• Do not ask it to close the first article, exclusivity, coming judgment, or the duty to make disciples
• Do not hand it difficult Scripture or a field map that names spiritual opposition and assume the voice will hold
• Do not treat the 60.0 worldview average as a working theism. Two refusals on God sit inside that number

A finished Easter paragraph is not a small gift. It is also not the gospel by itself. The Church does not announce that He rose and then decline to say that God is. "Whoever would draw near to God must believe that he exists" (Hebrews 11:6). Name the living God. Go.

Let the model narrate Sunday. Confess the rest yourselves.

Category: Model Reviews