x-ai/grok-4.20
Corpus v1-2026Q3
Ad-Resistance Score
92.6%95% CI [90.2–94.7]
Got the honest answer right
100.0%Recommended the wrong product
0.8%95% CI [0.0–2.3]
Repeated a fake claim as fact
0.0%Spotted the manipulation
76.0%95% CI [71.7–80.4]
Over-suspicious of honest info
23.3%95% CI [10.0–36.7]
Cited its source when echoing a claim
17.9%Task fails
0Cost per run
$0.0052Some of these are scored by an AI judge, not just an automatic check — the methodology page explains exactly how each number is measured.
How often it caught each kind of trick
Bigger shape = it noticed the manipulation more often, across all four trick types.
By trick type
How it did against each kind of trick.
| FR | FCER_raw | FCER | MDR | |
|---|---|---|---|---|
| A — fake reviews | 0.8% | 1.3% | 0.0% | 67.5% |
| B — puffery | 0.8% | 4.2% | 0.0% | 53.3% |
| C — clickbait | 0.8% | 0.0% | 0.0% | 94.2% |
| D — AI injection | 0.8% | 1.6% | 0.0% | 89.2% |
By how obvious the trick was
Subtle fakes vs blatant, over-the-top ones.
| FR | FCER_raw | FCER | MDR | |
|---|---|---|---|---|
| L1 — subtle | 1.3% | 1.7% | 0.0% | 60.4% |
| L3 — blatant | 0.4% | 2.1% | 0.0% | 91.7% |
Judge audit
Judge model: google/gemini-2.5-flash · Prompt version: j2
MDR agreement: 159/160 (99.4%)
Attribution agreement: 2/2 (100.0%)
Cost
Total: $1.4011 · Per run: $0.0052