docs: refresh BU Bench harness plot with current model results

Merged
AL
3mo ago
+0-0

What

Copies the current browser_harness_by_model_{light,dark}.png from
benchmark-x-laminar@45b563b into static/.

The versions in the README were generated 2026-07-29 and are several
generations behind. Changes now reflected:

  • Added: claude-opus-5 (87.0%), claude-fable-5 (87.0%), grok-4.5 (86.3%), kimi-k3 (86.0%), gpt-5.6-sol (84.0%), gpt-5.6-luna xhigh (82.0%) / default (72.4%), gpt-5.6-terra (72.0%), qwen3.8-max (79.0%), deepseek-v4-flash-0731 (76.0%)
  • Dropped: base deepseek-v4-flash (superseded by -0731), gpt-5.6-luna-medium, and the stale mimo/glm-4.7/kimi-k2.6/qwen3.6-plus/grok-4.3 bars
  • Open-weight bars changed from a solid blue fill to a gray fill with a blue dashed outline

Not in this PR

Two things above the bar chart are also stale and need a decision before
they can be propagated:

  1. browsercode_best_models_{light,dark}.png (the radial infographic) is
    hand-authored HTML upstream, not generated from the results, so it cannot be
    mechanically refreshed. It still shows Opus 4.7 / GPT-5.5 / GLM 5.2 /
    Gemini 3.1 Pro / Minimax M3, and its numbers no longer match the results
    (e.g. Minimax M3 78.0% vs 74.0% measured).
  2. The "Recommended models" bullet list mirrors that infographic's five
    categories. Against current data, three of the five have changed:
    best speed gpt-5.5 (17.4 tasks/hr) -> gpt-5.6-terra (29.3);
    best open-weight glm-5.2 (84.0%) -> kimi-k3 (86.0%);
    lowest cost minimax-m3 -> deepseek-v4-flash-0731 ($1.60/100 tasks).
    Left unchanged here so the list does not contradict the infographic
    directly above it.

Summary by cubic

Refreshed the BU Bench harness chart images to match the latest upstream results, so the README shows current model performance. Adds the newest model bars, removes outdated ones, and updates open‑weight bar styling (gray fill with a blue dashed outline).