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

Original Article ↗ Hacker News Discussion ↗

OpenAI’s new GPT‑5.6 Sol model is widely seen as a major step up in image and video understanding, especially for complex, real‑world tasks like UI critique, shopping assistance, and fine‑grained visual reasoning. However, benchmark results and user reports suggest Google’s Gemini 3.x family and Alibaba’s Qwen 3.8 often outperform it on object detection, counting, and cost‑efficiency, particularly for high‑volume or production workloads. Many commenters argue that traditional vision tools (e.g., OpenCV) and specialized models still beat general‑purpose LLMs on narrow tasks, but see frontier vision‑language models as increasingly valuable for auto‑labeling data, orchestrating tools, and handling messy multimodal inputs.

Overall sentiment on GPT‑5.6 Sol’s vision

  • Many find Sol a substantial step up over earlier OpenAI vision models, especially for:
    • Complex layouts (e.g., nested modals, app UIs).
    • Video captioning with sub‑second motion and robustness to FPS.
    • Hard OCR cases (faint reflections, tricky screenshots, sheet music).
  • Others report it still “as blind as any other model” for fine detail, visual “taste,” and strict design‑system adherence.
  • Vision is seen as clearly behind text capabilities; surprising failures (blank image → hallucinated kitchen, illegible dates, handwritten Cyrillic).

Comparisons: Gemini, Qwen, and others

  • Multiple commenters say Gemini (especially 3.7 Flash) is currently the best practical vision choice:
    • Beats Sol on most Roboflow benchmarks, particularly detection/counting.
    • Cheaper and fast enough for high‑volume tasks; Flash Lite praised for extraction, though flaky on large/complex inputs.
  • Some still favor Gemini Pro 3/3.1 for deep scientific reasoning, but 3.7 Flash is viewed as comparable or better for messy PDFs and vision tasks.
  • Qwen 3.8 (Max and 27B) gets strong reviews for vision and object detection; open‑weights 27B is highlighted for local use.
  • Other models mentioned: Fable for OCR, Luna and Seed Turbo 2.1 for strong vision descriptions, small models like minicpm‑v‑4.6 for on‑device use.

Benchmarks, methodology, and classical CV

  • Commenters spot ground‑truth and box-rotation issues in the pill/egg examples; image resolution vs EXIF orientation is debated.
  • Pill counting at ~80% accuracy is deemed far below production pharmacy needs.
  • Several argue these tasks are trivial with OpenCV/template matching; others counter that:
    • VLMs are valuable to auto‑annotate datasets.
    • The right pattern is LLMs orchestrating classical CV tools, not replacing them outright.

Design, UX, and “taste”

  • Mixed experiences on using LLMs to critique UI screenshots:
    • Some say Sol can decompose screens into coherent, reusable blocks.
    • Others find models inconsistent or cliched, especially with “AI‑looking” UIs.
  • Debate over how much of design/UX is objectively measurable vs “taste,” and whether models can capture those nuances.

Use cases and limitations

  • Reported successes: grocery‑aisle product finding and cropping, pharmacy pill counting assistance, diagram review, sheet music transcription.
  • Notable failures: path‑following puzzle without tools, crop disease misdiagnosis, depth/3D understanding (binocular benchmarks are asked for but not discussed).
  • Some see Sol and peers as impressive generalists but not yet reliable for safety‑critical or high‑precision production vision.