GPT-5.6 Sol is OpenAI's best vision model yet — Gemini still wins
Roboflow's benchmark says Sol's object detection leapt 3x to a usable 46.2 mAP, but at 2.5¢ an image Gemini's cheaper — plus a DuckDB 2.0 preview for builders.
Copy markdown
Detection went from broken to usable
Roboflow's new VLM benchmark puts GPT-5.6 Sol at 46.2 mAP@50 on object detection — up from GPT-5.5's 13.8, a 3.3x jump that turns a dead-end into something you can actually ship. Counting accuracy also climbed to 73%, from 64.9%.
But OCR slipped, and it's not cheap
Sol's OCR similarity dipped to 90.7% (below GPT-5.5's 91.2%) and text extraction fell to 82.5%. At ~2.5¢ and ~10s per image it's near the priciest option tested — Gemini 3.5 Flash matches or beats it on detection at ~0.8¢, roughly a third the cost.
Resize big images, or reach for Luna
Sol gets unstable above 2000×2000px at low reasoning effort, so downscale first. For cheap high-volume vision, the Luna variant runs under 0.5¢/image at ~5s — the best latency-to-cost trade in the GPT-5.6 lineup.
Elsewhere: DuckDB 2.0 preview is live
DuckDB 2.0 'Cyanoptera' preview builds (Aug 17) ship a stable client/server mode, async S3 I/O, a production VARIANT type for JSON-ish data, triggers, and ~40x faster recursive CTEs. The full release lands this fall.