$40 experiment: Opus 5 hits ~0.94 F1 on ExtractBench, Qwen3.8 matches at 1/3 the price

Ok-Challenge-7810 · reddit · 2026-08-25

An AI engineer with a computational physics background ran a $40 experiment on LlamaIndex's ExtractBench (schema-guided document extraction): 36 government docs, customs forms and municipal audits, text pulled with pypdf, one-shot extraction with Claude Opus 5, Qwen3.8-2.4T and Qwen3.6-35B via OpenRouter, scored with the benchmark's cell-level F1.

Two surprises: Opus 5 one-shot on plain text scored 0.94, on par with the coding agents in the paper at a fraction of the per-page cost; and Qwen3.8 basically matched it (0.936) at a third of the price. F1 also correlated much more with document length (ρ ≈ −0.50) than with the number of values to extract (ρ ≈ −0.28), fitting the recent pre-inference routing paper (arXiv 2608.06607).

Caveats: n=36, short/medium docs only, text-layer only (scanned forms excluded), and the scorer is a reimplementation, so absolute numbers may drift from the leaderboard. Notebook is public.

Original post →

More from Models

Models channel →