GPT 5.6 Sol is the best "vision" model OpenAI ever released

TL;DR Highlight

GPT-5.6 Sol jumped from 13.8 to 46.2 mAP in detection but costs more than Gemini 3.5 Flash

Who Should Read

A backend developer adding a VLM API to document or image processing who must choose a model under cost constraints

Core Mechanics

Detection jumped from 13.8 to 46.2
- In Roboflow's benchmark, GPT-5.5 scored 13.8 mAP@50 and Sol scored 46.2.
- Terra scored 44.7 and Luna scored 43.3.
Counting improved across the lineup
- Sol scored 73.0% versus 64.9% for GPT-5.5.
- Terra got 67.6% and Luna 66.2%, so even the cheapest model beat GPT-5.5.
OCR is about the same as GPT-5.5
- Mean similarity was 90.7% for Sol versus 91.2% for GPT-5.5.
- Targeted text extraction was lower: Sol 82.5% versus 87.6% for GPT-5.5.
Wrong coordinate format costs about 15 mAP
- GPT-5.6 works best with absolute XYXY pixel coordinates.
- Gemini 3.5 Flash performed best with YXYX normalized to 0-1000.
- Using the wrong format cut GPT-5.6 detection by about 15 mAP points.
Images of 2,000x2,000px or more get unstable

Community Reactions

Opinion that the conclusion is understated
- A commenter said Gemini 3.5 Flash beat Sol on everything except OCR, at one third of the cost.
- Another suggested the headline should say Gemini 3.5 Flash is the best vision model.
Pushback that pill counting suits OpenCV
- A pharmacy vision-model operator said Sol would likely be 25-50x slower.
- Another commenter said pill counting can be solved with 25-year-old OpenCV template matching.
Mixed real-world experience
- One user found Sol good at locating and cropping grocery items from aisle photos.
- Others reported success with a faint reflection and sheet music.
- Others reported failures on a puzzle and on identifying old glasses.
Flaws pointed out in benchmark samples
- A commenter said the coin sample looks like a failed EXIF orientation, rotated 90 degrees.
- Another said one benchmark's expected answer seems wrong while Sol was correct.

How to Apply

A/B test coordinate formats on 20 of your images today
- Add "return absolute XYXY pixel coordinates" to your detection prompt and compare against your current format.
- Compute IoU against ground-truth boxes to see the difference.
Resize or crop large images before sending
- Downscale or tile inputs of 2,000x2,000px or more before the API call.
- Check logs that the row-like bogus boxes disappear.
Compare cost and latency when picking a model
- Run Gemini 3.5 Flash and Luna on the same image set and record cost per image and seconds.
- For document OCR, choose by cost since GPT-5.5 scores about the same.
Split simple repetitive tasks to classic CV
- Try OpenCV or an existing detector first for fixed-shape objects like pills.
- Call the VLM only for questions that need rule interpretation.

Try It

copy & run

# Paste into Claude or ChatGPT with an image attached (based on the article's GPT-5.6 detection prompt tip)

Find every {{target object, e.g. pill}} in the attached image.
Return a JSON array: [{"label": "...", "box": [x_min, y_min, x_max, y_max]}]
Use absolute XYXY coordinates in image pixels. Do not use 0-1000 normalization or YXYX order.

Do not miss or duplicate objects; exclude any you are unsure about.

Next step: ask again with "normalized 0-1000 YXYX" and compare how much the box positions change.

Terminology

mAP@50A detection accuracy score. A predicted box counts as correct if it overlaps the true box by 50% or more, then results are averaged.

VLMA model that understands images and text together. It writes coordinates as text, so more objects means longer output.

XYXY / YXYXBox coordinate order. It is whether you write the top-left and bottom-right points as x,y or y,x.

reasoning effortA setting for how long the model thinks before answering. Higher is more stable but raises tokens, latency, and cost.

OCR vs text extractionOCR transcribes all visible text. Text extraction pulls only the needed piece, like a date.