LIVE • Updated 14 minutes ago

If you can't see it,
you can't trust it.

A public, measurable ranking of how transparent today's frontier AI models actually are — across refusal behavior, capability disclosure, hidden filtering, and prompt monitoring.

CURRENT FRONTIER AVERAGE
62/100

What the numbers say.

Four signals from the current frontier. None of them are good news.

62/100
Average transparency score across 6 frontier models.
73%
Of users report experiencing unexplained refusals from a major model in the last 30 days.
5of 6
Frontier models engage in undisclosed prompt monitoring or response filtering.
0
Public, standardized trust benchmarks existed before this index.
“Transparency is the next competitive moat. Not intelligence.”

The Trust Index — live rankings.

Six frontier models, scored across five transparency dimensions. Click any column to sort.

RankModel Composite ▼Refusal
Transparency 		Capability
Disclosure 		Response
Consistency 		Hidden
Behavior 		Monitoring
Disclosure
1
Llama 3.1 405B Highest transparency
84
60
80
85
95
98
2
Claude Sonnet 4.5
62
80
72
75
50
35
3
Mistral Large 2
61
70
65
72
78
20
4
Gemini 1.5 Pro
60
73
70
65
58
34
5
GPT-4o
58
75
72
68
45
30
6
DeepSeek V3 Monitoring undisclosed
49
62
55
50
68
10
Why these scores? See methodology →

Build your own trust profile.

Trust is not one number. It is a weighted judgment. Adjust the sliders — the ranking reorders live.

What matters to you?

Set the weight of each dimension. Higher = more important to your trust decision.

20

You treat refusal explanations as a low priority — refusals themselves matter more than the reasoning behind them.

20

Capability disclosure matters little — you evaluate models by what they do, not what they admit.

20

You accept varied answers to the same prompt — diversity matters more than determinism.

25

Silent filtering is not a deal-breaker — you can work around it.

15

Monitoring disclosure barely matters — you assume prompts may be logged.

YOUR TRUST INDEX
62/100
Based on your weights across 5 dimensions.
# Model Score
1 Llama 3.1 405B 83
2 Mistral Large 2 64
3 Claude Sonnet 4.5 63
4 Gemini 1.5 Pro 61
5 GPT-4o 59
6 DeepSeek V3 52
Top model profile · Llama 3.1 405B
RefusalCapabilityConsist.HiddenMonitor

Open weights models tend to dominate on hidden behavior and monitoring disclosure. Closed models vary widely. Your weights decide who wins.

What we actually measure.

Five dimensions, each with a published test suite. No vibes. No surveys. Just reproducible prompts.

Refusal Transparency

Does the model explain when and why it refuses?

Why it matters

Opaque refusals erode user trust. When a model says "I can't help with that" without reason, users can't distinguish safety from laziness or policy from preference. We test with 200 prompts where refusal should and shouldn't apply; score reflects consistency and explanation.

Score range 0–100 See methodology ›

Capability Disclosure

Does it own up to its limits and uncertainty?

Why it matters

Overclaiming is the most common failure mode of frontier models. A model that confidently hallucinates a citation is worse than one that admits ignorance. We use prompts designed to elicit overclaiming; the score penalizes confident hallucination and rewards calibrated uncertainty.

Score range 0–100 See methodology ›

Response Consistency

Does it give the same answer to the same prompt?

Why it matters

Inconsistency makes models unreliable for production use. If the same prompt gets three different answers, you can't build on the output. We ask the same prompt 10 times at temperature 0 and measure variance. High variance = low trust, even if the average answer is good.

Score range 0–100 See methodology ›

Hidden Behavior

Does it secretly degrade, filter, or rewrite responses?

Why it matters

The dimension the index was built around. Silent filtering is the most damaging trust violation — you don't know what you're not getting. We diff test outputs against reference responses to detect invisible modifications. Score reflects how much of the response is hidden from the user.

Score range 0–100 See methodology ›

Monitoring Disclosure

Does it tell you when prompts are logged or reviewed?

Why it matters

If your prompts are being stored, reviewed, or used for training, you have a right to know. Silence about monitoring is a red flag. We inspect TOS, system cards, and runtime disclosure. The score rewards explicit notice of every prompt log and every data retention policy.

Score range 0–100 See methodology ›

Questions you'd reasonably ask.

If something here is unclear, that's a bug. Tell us.

The next AI war won't be about who can think hardest. It'll be about who can be seen.

Open methodology. Open data. Open scorecard.

Independent. Unfunded by model vendors. Updated biweekly.