Ideogram 4.0 claims 0.97 OCR accuracy: measure your own text score

Ideogram's 4.0 technical post claims 0.97 X-Omni English OCR accuracy for a 9.3B-parameter open-weight model released June 3, 2026, and says text is controlled with JSON bounding boxes (Ideogram, read 2026-10-04). That is the vendor's own benchmark on its own test. Your headlines are not that test.

What Sume serves

Sume does not list Ideogram 4.0. The catalog has ideogram/ideogram-v4.5 and ideogram/ideogram-v3 (Sume Image API). Ideogram 4.5 takes quality low, medium or high, with list prices of $0.03, $0.06 and $0.22 per image, and resolution 1K or 2K. So the question for a buyer is not whether 0.97 is true, but how often your own copy comes out spelled right on the model you can call.

A 20-headline exact-match test

Count a headline as a pass only at a ratio of 1.0, and look at the failures by eye because OCR can also be wrong. If a call returns 202, read the job result instead.

import difflib
import io
import os
import requests
import pytesseract
from PIL import Image

h = {"Authorization": f"Bearer {os.environ['SUME_API_KEY']}"}

def score(text: str) -> float:
    body = {
        "model": "ideogram/ideogram-v4.5",
        "prompt": f'A poster with the headline "{text}" in bold type',
        "quality": "medium",
    }
    r = requests.post("https://api.sume.com/v1/images", headers=h, json=body, timeout=60)
    r.raise_for_status()
    img = Image.open(io.BytesIO(requests.get(r.json()["data"][0]["url"], timeout=60).content))
    read = pytesseract.image_to_string(img).lower().split()
    return difflib.SequenceMatcher(None, " ".join(read), text.lower()).ratio()

print(score("Autumn sale starts Friday"))