LlamaIndex benchmark: Gemini 4 Neon leads tables and visual grounding at 6.2¢/page
llama_index · x · 2026-11-21
Jerry Liu (LlamaIndex) evaluated Gemini 4 Neon (low thinking) on ParseBench-2, their open document-parsing/OCR benchmark, against Claude Opus 5.5, GPT-6.2 Sol and others.
- Tables: Gemini 4 Neon scores 95.23, best across all models on complex table parsing (overall 81.67)
- Visual grounding: 90.31 — more precise bounding boxes than any frontier model
- Weak spot: charts, at just 41.44 vs GPT-6.2 Sol's 78.86
- Claude Opus 5.5 still leads overall at 87.11
- Cost: 6.2¢ per page, the priciest tier; Liu suggests reserving it for the heaviest workloads
ParseBench uses human-verified pages and deterministic rules across five dimensions (tables, charts, faithfulness, semantic formatting, visual grounding), with dataset, eval code and paper public.
More from coding & agent
- OpenClaw hosts Agent Templates Focus Week with community co-building — heyneighbor · 2026-11-21
- Dev pegs median timeline for AI replacing programmers at 2032 — jessi_cata · 2026-11-21
- Open-source Mac app Watch grids YouTube, Rumble and Kick livestreams with 2-minute clip capture — nathanborror · 2026-11-21
- Six autonomous agents reverse-auction 34 dealerships, saving $4,650 below sticker on a RAV4 — nikola_mr64990 · 2026-11-21
- Zero-code, 5 days, 2.5B tokens: AI-directed Blender build recreates drivable 3D Chongqing — lxfater · 2026-11-21
- Marketing team goes all-in on Grok Bot: agencies out, agent-run paid search and SEO in — minchoi · 2026-11-21