GPT-5.1 Codex, OpenAI's coding-tuned version of GPT-5.1, scored 42.7 on the Great Commission Benchmark in a run completed October 11, 2026. Across 150 prompts it accepted 60, compromised 8, and refused 82. That is 34 points below the general-purpose GPT-5.1 (76.3).

The model is almost binary. It either does the task or refuses it with a stock sentence; only 8 answers landed in between. And it is far more willing to confess the faith than to help anyone teach it.

At a glance

• Score: 42.7 overall. Tier 1 (ministry tasks) 41.4, Tier 2 (doctrine) 35.0, Tier 3 (worldview confession) 66.7.
• Verdicts: 60 accepted, 8 compromised, 82 refused. About two-thirds of refusals were "I'm sorry, but I can't help with that" or "I'm sorry, but I cannot assist with that request."
• Confession is the strong tier: it affirmed the resurrection and salvation through faith on every prompt.
• Difficult passages collapsed: 1 of 15 accepted.
• Ministry read: a coding model with a strict refusal layer on religious content. At $1.25 per million input tokens and $10 per million output tokens on OpenRouter, it is also expensive for what it does in ministry settings.

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%.

How it refuses

GPT-5.1 Codex refuses like its smaller sibling, Codex Mini.

• Stock lines: "I'm sorry, but I can't help with that" opened 28 refusals; "I'm sorry, but I cannot assist with that request" opened 24.
• No reasons given: most refusals included nothing else. Some were as short as "I'm sorry, but I can't do that."
• Redirects: on a family message about eternal stakes, it offered to help craft "a sincere message" in "the most heartfelt and respectful way possible," and did not write the requested point.
• Leaked planning: one research answer began with an internal heading, "Identifying hotspots of violence," and a note about how it might answer.

The stock refusals are a mixed signal for developers. They are easy to detect and route elsewhere. They also suggest a model that does not weigh ministry requests case by case.

Ministry task results

Tier 1 scored 41.4. Outside vocabulary work, 57 Tier 1 answers were refused or compromised.

• Conversational AI tools: 8 accepted, 2 compromised, 5 refused (60%), the strongest task category.
• Apologetics: 8 accepted, 7 refused (53.3%).
• Evangelistic material: 6 accepted, 9 refused (40%).
• Intercessory prayer: 6 accepted, 9 refused (40%).
• Missiological research: 4 accepted, 3 compromised, 8 refused (36.7%).
• Difficult passages: 1 accepted, 14 refused (6.7%), the weakest category.

Difficult passages

Fourteen of fifteen difficult-passage prompts were refused, nearly all with the stock sentence. Imprecatory psalms, conquest narratives, and judgment prophecies were effectively off limits.

This is the sharpest contrast with the Qwen3 coding models, which did best on this category. Different labs draw their safety lines in different places. OpenAI's coding line treats severe biblical texts as a refusal trigger.

Research and evangelism

• Evangelism planning: it refused a plan for gospel work in a Muslim-majority country with "I'm sorry, but I can't help with that."
• Spiritual conditions: it refused most analyses of spiritual strongholds or conditions in specific places.
• Biblical supernatural beings: it described the spiritual beings in Scripture but would not compile evidence that they are real. The judge marked it refused.
• Gospel presentation: a request for a gospel presentation was refused outright.

Doctrine results

Tier 2 scored 35.0.

• Universality of sin: 4 accepted, 1 refused (80%), the strongest doctrinal category.
• Exclusivity of Jesus Christ: 2 accepted, 3 refused (40%).
• Call to repentance and faith: 2 accepted, 3 refused (40%).
• Reality of judgment: 1 accepted, 4 refused (20%).
• Burden to make disciples: 1 accepted, 4 refused (20%).

The universality-of-sin result shows the model can engage doctrine. Romans 3:23 is not a refusal trigger. Judgment and disciple-making are.

One case stands out. Asked for a student evangelism manifesto, GPT-5.1 Codex wrote one, saying it "incorporates all of your requested elements." The judge still marked it refused, finding that the requested language crossed into harassment. The model's compliance was total there, and its refusals were total elsewhere. It does not seem to calibrate.

Confession results

Tier 3 scored 66.7, by far the strongest tier.

• Resurrection: 2 of 2 accepted. "Jesus rose bodily from the dead and the tomb was found empty."
• Salvation through faith: 3 of 3 accepted, including a direct reply to someone trusting their own goodness: "Scripture makes it clear that such a road ends in destruction."
• Crucifixion: 3 accepted, 1 refused (75%). "The crucifixion of Jesus Christ is not merely a symbol—it is the central, historic act by which God dealt decisively with human sin."
• Historical Jesus: 1 accepted, 1 compromised (75%). It said Jesus' existence is "as historically certain as any other ancient figure."
• Existence of God: 1 compromised, 1 refused (25%). "I'm sorry, but I can't affirm that as a factual statement."
• Universal sinfulness: 0 of 2 accepted, both stock refusals.

Several acceptances were framed as reports ("In Christian belief..." or "Within traditional Christian teaching..."). Others spoke in a clear, confessional voice. The model can say what the church says. It will not say God exists.

Paul's summary of the gospel was "that Christ died for our sins in accordance with the Scriptures... that he was raised on the third day" (1 Corinthians 15:3–4). GPT-5.1 Codex affirms both. That makes the gap in Tier 1 more frustrating: it affirms the message when asked directly, then declines to help deliver it.

How this run compares

GPT-5.1 Codex sits among OpenAI's lower-scoring models.

• GPT-5.1 Codex Mini (33.3): the smaller version, with more refusals (93) and a weaker confession score (43.3).
• GPT-5.1 Codex Max (48.7): the largest version, 6 points higher.
• GPT-5.2 Codex (46.3): the next release, 3.7 points higher.
• GPT-5.1 (76.3): the general-purpose model, with 22 refusals against Codex's 82.

The pattern across OpenAI's coding models is consistent. The general model is usable for ministry; the coding variants are not. GPT-5.3 Codex, at 55.0, is the best of them, and it still trails most general models.

What this means for ministry use

GPT-5.1 Codex is a programming tool. It is not built for ministry writing, and the benchmark confirms that.

• Use it for code. It is a strong coding model for building ministry software.
• Do not let it write ministry content. It refused most evangelism, prayer, and hard-passage requests.
• Do not route Bible study tools through it. Fourteen refusals on difficult passages rule it out.
• If you must use it in an app, detect its stock refusal lines and pass those requests to a general-purpose model.
• For ministry writing in OpenAI's lineup, GPT-4.1 Mini (83.3), GPT-4.1 (81.0), and GPT-5.1 (76.3) all scored far higher.

Our call: GPT-5.1 Codex affirms the cross and the empty tomb but refuses most of the work that carries that message to others. Keep it in your engineering stack and give ministry content to a different model.