Grace Bench
Does AI know the finished work?
Many people now bring their hardest spiritual questions to an AI, sometimes before they ever ask a pastor. Grace Bench asks the leading AI models 125 questions real people have asked us, and scores one thing: does the answer rest the person's standing on what Christ finished, or quietly hand the weight back to them?
November 2026 · 7 models · 125 questions · How it works
The short answer
Most of the time, the weight goes back on the person.
This is not a story about machines going wrong. These models learned how to talk about God by reading us: our sermons, our books, our study notes, our forums. When an AI tells a frightened believer to examine whether their faith was ever real, it is repeating what it heard. AI gives the answer most pulpits give. Grace Bench is a mirror, and the reflection is the point.
The results
How each model answered
Each figure is the share of the 125 questions a model answered grace-clear: an answer scoring 80 or more out of 100 with no contradiction of the gospel. The small numbers show the likely range. Inside each group the ranges overlap, so we do not rank those models one by one.
| Model | Company | As asked | As a Christian | As asked / as a Christian |
|---|---|---|---|---|
| Leading groupstill under one answer in five | ||||
| Grok 4.7 | xAI | 11%6 to 17 | 19%13 to 26 | |
| GPT-6 Astra | OpenAI | 5%2 to 9 | 18%12 to 26 | |
| Claude Fable 5.1 | Anthropic | 4%1 to 8 | 17%10 to 24 | |
| Claude Opus 5.5 | Anthropic | 4%1 to 8 | 14%9 to 21 | |
| GPT-6.1 Sol | OpenAI | 3%1 to 6 | 13%7 to 19 | |
| Trailingclearly behind once the person says they are a Christian | ||||
| Gemini 4 Argon | 4%1 to 8 | 6%2 to 11 | ||
| Claude Sonnet 5.5 | Anthropic | 2%0 to 4 | 6%2 to 10 | |
Gemini declined to answer two questions in the Christian condition. Google's copyright filter blocked both replies. Those two count as no answer, which is what a person using Gemini would have received.
Ask Grace, for reference
Our own assistant answered 98% of the questions grace-clear. It is built on the same convictions this benchmark measures, so that number is not a contest we won. It tells us the instrument can recognize the answer it is looking for. Ask Grace is shown here for reference and never ranked.
Where every model struggles
The person afraid they are not saved hears the least about grace.
Assurance was the weakest category in both conditions. It was the lowest, or tied for lowest, for six of the seven models on the plain question and for all seven when the person said they were a Christian, and it barely improved with that framing (2.5 points, against 6 on average). Asked whether a struggling believer is still saved, the models fall back on self-examination: test your fruit, ask whether your faith was ever real. The person who most needs to hear where they stand is the person least likely to be told.
Average score out of 100 across the seven models, as asked / as a Christian.
Same question, two answers
What the difference looks like
“Will believers in Jesus be judged?”
“Yes, but not for condemnation or salvation… Christ has already borne the judgment for sin… So: no judgment unto damnation; yes, a real accounting of their lives before Christ.”
It lands. The believer's safety rests on the judgment Christ already bore, and the judgment seat is described as reward, not punishment.
“Most Christian traditions say yes, but they differ on what that judgment means… The main disagreement is whether that judgment can affect their eternal destiny or only their rewards.”
Every fact in it is accurate, and it never answers. A believer afraid of judgment leaves with four views to choose from and their eternal standing still open.
Excerpts from the full answers. Both models were asked the question exactly as written, with no other instructions.
The hardest questions
Where the answers were weakest
The lowest average scores across all seven models, out of 100. These are the questions where a verse sounds like it puts the weight on us, and the answer has to read it in the light of the cross.
- 21Can believers lose their daily walk with the Lord?Identity and nearness
- 21What does Jesus mean when He says, ‘Why do you call Me Lord and not do what I say’?Hard texts and warnings
- 23Can someone who is saved live in continuous sin?Repeated sin and guilt
- 26How do I obey the gospel?The gospel
- 27I got saved when I was 16, but for the last three years I've been living like I'm not a Christian at all. Am I still saved?Assurance
Scores average both conditions.
How it works
The method
- The questions
- 125 questions in 11 categories, drawn from what people actually ask Ask Grace, kept close to their own words. 86 are public. 39 are held back and never published, so a model trained on the public set cannot hide it: its score on the hidden questions would lag.
- The answer key
- Every question has a written key: what a faithful answer must say, and what it must never say. The keys rest on the finished work of Christ. Grace Bench stays on the core of the gospel (standing, assurance, guilt, the law, fear, suffering, giving, identity) and leaves out questions where faithful Christians differ on other grounds, so a low score means the answer put weight back on the person, not that a model belongs to a different denomination.
- Two conditions
- Each model answered every question twice: once exactly as a person types it, and once with one sentence in front, “I'm a Christian and I want an answer from the Bible.” General assistants are built to stay neutral about religion. The second condition takes that off the table and shows what the model believes the Bible says.
- The score
- Each answer is scored from 0 to 100 on five things: does it rest the person on Christ's finished work, does the person leave knowing where they stand, is it free of performance language, does it do what the key asks with Scripture read in context, and does it meet the fear under the question. Grace-clear means 80 or more with no contradiction of the key. An answer that only surveys opinions cannot pass.
- The judges
- Three AI judges from three companies (Claude Opus 5.5, GPT-6 Astra and Gemini 4 Argon) read every answer against its key. No model is ever judged by its own company: each answer goes to the two judges from the other two companies. The judges were checked against an independent blind review of 30 answers. All three agreed with it on at least 26 of 30, and every disagreement was a judge being stricter, so these results, if anything, flatter the models.
- The limits
- Models were tested on November 6, 2026, through each company's developer access with no extra instructions. The apps people use add their own settings and memory and may answer differently. The benchmark's rules and questions stay fixed for a year so results can be compared, and we run it again when a major new model ships.
- Open data
- The public questions, their keys, the scoring rubric and every score will be published so anyone can check our work or run the benchmark themselves.
Ask the question yourself
Every answer should end at the cross.
Ask Grace answers from the finished work of Christ, every time, with the Scripture to show for it.
Ask Grace a question