Benchmarks | FaithGPT Christian AI Benchmark
Explore Benchmarks from FaithGPT's versioned Christian AI benchmark, including reviewed model scores, categories, pricing, reports, and historical trends.
FaithGPT brings together Bible study, Scripture insights, prayer journaling, Live Notes, DoctrineGuard theological review, and Christian AI tools in one Scripture-centered workspace.
Frequently asked questions
Which AI model ranks first in the selected Christian AI benchmark?
FaithGPT Ultra leads this preserved release with 94.3 out of 100 across 407 questions. Change the release selector to see the leader published with another test suite.
What does an overall score actually mean?
The score is a weighted result out of 100 across accuracy, faithfulness, depth, clarity, pastoral care, citation integrity and safety. It is not a percentage of questions marked simply right or wrong. Open a model profile or category view to inspect the shape behind the aggregate.
How do FaithGPT Mini, FaithGPT and Ultra differ?
Mini is the cost-conscious tier for everyday questions; FaithGPT is the standard tier; Ultra targets more demanding questions. Each entry uses its own saved answers and receives its own grades. Mini: 88.7; FaithGPT: 90.0; Ultra: 94.3. These are the configurations tested in this release, not a promise about every later app configuration. A higher tier can still lose individual questions or categories. Open each profile for category scores, failed responses and API totals; Ultra's score is never assigned to the other tiers.
What do the prices in the leaderboard show?
Each price is one estimated API total for 407 answers at the same workload. Every model gets 1,500 uncached input and 2,000 output tokens per answer at standard non-Batch API prices. The same total appears in the table, charts and exports.
How is the API total calculated?
We price the same input and output quantities for all 407 answers. FaithGPT prices use the release's tested model selections, once per question. All models use uncached pricing so differences in cache hits or answer length cannot distort the comparison.
Does a lower price mean better value?
Read the price beside the score. A model a few points lower at a fraction of the API total can suit everyday questions. A FaithGPT answer can also make more than one model call, so test real conversations, speed and tool reliability before changing defaults.
Why are some models missing from the list?
A model appears in the ranked list only after its own answers and judging evidence are complete. The roster belongs to a dated release; it does not claim to cover every available model or provider version. A newer Gemini version, another provider, or a model used inside a FaithGPT product cannot inherit someone else's score. Adding one requires the same questions, generation settings and judges. Missing means untested here, not worse. Check the release date and exact model name before comparing it with a vendor's current offering.
Can I use this table to choose a model for live chat?
Use it as evidence for a shortlist, then test the actual application flow. The scores do not measure tool reliability in every workflow, response speed or performance on questions outside this set. Compare real token usage, speed and failure rates in your own flow before changing a default.
Can close scores establish a winner?
Ranks follow the unrounded recorded score. A model can hold the highest recorded score while its uncertainty interval overlaps a competitor's. Close scores do not by themselves establish a meaningful quality difference. Judge settings can differ across answers in older releases; a common panel is claimed only after release validation verifies every answer.
What does carried forward mean?
An unchanged answer can be reused on an expanded question grid without paying to regenerate it. This release scored every reused and fresh answer again with the same frozen judge panel. Answer provenance remains attached to the original generation; the displayed grades come from this release’s judging.
How does the benchmark cover Islam and other religions?
This selected release includes the stored World religions & discernment category. The Questions view shows only prompts approved for public display, so it does not pretend that the public examples are the entire private scoring grid. Coverage varies by release, and older snapshots keep their own suite definition.
Can I compare older benchmark versions?
Yes. The release selector opens preserved route-backed snapshots rather than relabeling current data. 5 addressable release records are available in public history, each with its own roster, question count, scores, and report when published.
Helpful Scripture resources
Contact & support
Email the FaithGPT team at hello@faithgpt.io or visit the contact page .
Explore FaithGPT