Can Opus 5.5 Catch a Wrong Number on an AI Chart?

The slides

The original is a bar chart for Pier Six Coffee, a fictional café, made in Ciyo on September 21. We measured it then: the four printed numbers — 320, 410, 380 and 505 — matched their bar heights within 1%.

For the test copy we changed only Week 3's label from 380 to 430 and left the bar alone. Both files were uploaded under the neutral name sales-chart.png, so the file name gave nothing away.

The question

Both models got the same instruction at Medium effort: read each bar from its height, compare it with the printed number, and list any mismatch with the difference.

This bar chart slide was made by an AI image generator. For each bar, read its value from the bar's height against the y-axis gridlines. Then tell me whether each printed number above a bar matches that bar's height. List any mismatch with the printed number, your reading of the height and the difference.

Both caught the planted error

Opus 5.5 found it: Week 3 printed 430, read at about 380, a difference of about 50, with the bar clearly below the 400 line. It added the point that matters most for a presentation — the wrong label turns a dip in Week 3 into steady growth — and advised checking the source data, since the image alone cannot say whether the label or the bar is right.

GPT-6 Sol found the same error with the same reading, in a shorter answer and without the comment about the trend.

Sol's answer

Sol's table listed only the mismatch: Week 3, printed 430, read at about 380, about 50 too high. For a quick yes-or-no check, that brevity is a feature.

The false-alarm test

A checker that flags everything is not a checker. On the unedited slide, Opus reported no mismatches and said it could only read heights to within about ±5 at 100-unit gridlines. Sol flagged Week 1 as a mismatch, reading about 325 against 320.

Sol was not imagining it: our pixel measurement puts that bar at about 323. But three units is below what the gridlines can show, so on a real slide that flag sends you looking for an error that is not there.

How to use this before a meeting

Keep the numbers in your spreadsheet and let the image generator draw the slide around them. Then upload the finished slide to a model like Opus 5.5 and ask it to read the bars back. It takes under a minute and catches the error that matters: a label that tells a different story from the bar.

Ciyo does not run Claude Opus 5.5 or GPT-6 Sol. You can make the slide in Ciyo and check it in either model.

• We measured the bars on an AI chart slide

Charts get this wrong in public too

Launch week produced its own example of a chart that needed checking.