How accurate is it?

Every number on this page was measured on coins certified by PCGS or NGC that the grader never saw while it was trained, using the same model and settings this site runs today. Where it is weak, we say so.

Model v3.11 + v3.12, measured 18 October 2026 (seller-photo ranges updated 19 October). We re-measure with every model release.

65%

of 5,912 auction-photo coins graded within one point of the slab

83%

of those graded MS65 to MS70 within one point

91%

of slab grades fell inside the 90% range we showed, as a 90% range should

91%

how often PCGS and NGC agree within one point on the same mint-state coin: the ceiling

Auction photographs

Both sides of the coin, photographed in a studio, the way major auction houses show them. This is the grader at its best.

Slab grade	Coins	Average error	Within 1 point	Within 3 points	Within 1 grade step	Inside our 90% range
All coins	5,912	1.3	65%	90%	89%	91%
Circulated (1–49)	753	3.7	22%	54%	83%	90%
About uncirculated (50–58)	786	2.4	31%	72%	81%	93%
Mint state 60–64	1,692	0.8	72%	98%	87%	89%
Mint state 65–70	2,681	0.6	83%	99.7%	94%	92%

Seller photographs

One photo per coin, taken by eBay sellers through the holder. Closer to what you will point the extension at, and clearly harder. These are modern US coins, many of them proofs, so they are not every kind of eBay listing.

Slab grade	Coins	Average error	Within 1 point	Within 3 points	Within 1 grade step	Inside our 90% range
All coins	1,427	3.0	40%	74%	61%	87%
Circulated (1–49)	157	8.5	13%	29%	51%	78%
About uncirculated (50–58)	240	4.6	21%	45%	48%	84%
Mint state 60–64	252	2.2	39%	81%	54%	82%
Mint state 65–70	778	1.6	52%	90%	70%	91%

Raw coins photographed out of the holder (506 PCGS PhotoGrade reference coins): average error 2.7 points, 69% within three points, 87% inside our 90% range.

“Within one grade” versus “within one point”

A grading step is not always a point. From MS60 up the steps are one point apart, but VF20 to VF25 is one step and five points. We report both. Measured in steps, the grader lands within one step of the slab on 89% of auction photos and 61% of seller photos.

For scale: when PCGS and NGC graded the same mint-state coin (591 coins we could match across resales), they gave the same grade 52% of the time and agreed within one point 91% of the time. A claim of 95% or more “within one grade” would mean agreeing with one grading service more closely than the two services agree with each other. Ask how such a number was measured.

When we hedge or decline

The grader refuses photos it cannot read (too small, blurry, glare, not a coin) before you spend a credit, and it says “at best” with a wider range when a coin or photo is outside what it knows well.

Auction photos: 1.5% declined, 11% hedged. The “at best” ceiling was right 75% of the time.
Seller photos: 0.1% declined, 46% hedged. The ceiling was right 63% of the time, and only 26% of the time on coins slabbed MS65 or better.

Where it is weak

Worn modern coins
Designs we rarely saw worn in training (Eisenhower and Kennedy, clad Washington and state quarters, Jefferson, Roosevelt, Memorial cents, small dollars) read about 10 points too high in seller photos when circulated: average error 11.6 on 64 coins. We warn on most of them, and the warning is still not enough.
Circulated coins in seller photos
Average error 8.5 points on 157 coins graded below 50, reading 5 points high on average. On auction photos the same grades average 3.7.
Our “at best” warnings on top-grade coins
When we say “at best MS64” on a seller photo of a coin slabbed MS65 or better, we are usually wrong: the warning was right on only 26% of those 186 coins. These coins read low, not high.
World coins and tokens
None were in training. On live auction lots, sharp two-side photos of Mexican coins came within 1.5 points on average (78 coins), but holder photos of Civil War tokens and some 8 reales read 3 to 10 points low. We are collecting these coins now.
Proofs
The number is usable (average error 0.7 on auction photos, 1.5 on seller photos), but the result is labelled on the mint-state scale: we do not yet tell you it is a proof.
Problem coins
Cleaned, tooled or damaged coins are graded as if they were not. Of 51 “details” coins we met on live auction sites, none was flagged.
Soft, small or glary photos
Soft seller photos average 4.7 points of error against 2.7 for sharp ones; glare reads about 1.7 points high. The photo check warns, but it declines very few.

How we measured

The grade compared is our expected grade, the number on the badge, against the grade on the certification label. Average error is the average distance between them in points.
Held out: the auction coins were set aside by identifier before training, and any coin we could match to a resale of a training coin was removed. The seller-photo coins were set aside by listing; no listing or photo is in both sets.
Trained on: About 657,000 certified coins: 653,000 from the Heritage Auctions archive (US coins from 1652 to 2019, both sides, studio photographs), 2,154 PCGS PhotoGrade reference coins, 309 PCGS registry coins and 1,439 certified eBay listings photographed by their sellers. No world coins. Coins with a problem (details) grade are left out.
As shown: every figure includes the warnings and wider ranges the site adds, because that is what you see.

Spotted a grade that looks wrong? Tell us: those reports are how we find the weak spots above.