Variant B — the annotated staircase

Design thesis: a Pareto frontier is a staircase, not a scatterplot — draw the dominant steps in plain SVG, label every step directly with the offering and its exact prices, drop a ghost dot for every dominated model with a thin arrow to the step it fails to reach, and place one editorial callout per chart naming the trade the data actually argues for.

Solid dots + stepped line = frontier · gold dot = dominated within 10% of the step it shadows · gray dots = dominated (arrow points at the frontier step that dominates them) · † = UNVERIFIED modeled cap. Hover a point to see exact values.

Sources: AA Intelligence Index v4.1.1 (current scale only — older articles use renormalized scales and are never mixed in) · GPQA Diamond AA-run · SWE-bench Verified via vals.ai independent same-harness (Mini-SWE-agent); openlm.ai rows are aggregator/vendor-reported · prices: OpenCode Go + ClinePass published tables and the live OpenRouter catalog · workload shapes: agent-chat from real 30-day traces, coding-session from OpenCode Go published request patterns. † caps are UNVERIFIED models: ClinePass assumed $35; $20-tier subscriptions modeled at 8× purchase price ($160) per community reporting. Y-axis labels follow each frontier's per-row secondary_label (note: math-smarts-coding's axes string says SWE-bench-Verified but its rows carry GPQA-Diamond values — the label shown is the authoritative one). Models without verified scores are excluded-with-reason (muse-glimmer; GLM-5.3/Flash and muse-spark lack verifiable GPQA — notable because GLM-5.3 leads the general index). ≈-near is computed client-side as: within 10% of the nearest strictly-better-or-tie row on the frontier's primary axis (cost or speed); it reproduces the generator's (+X% off) chips in 13 of 14 cases — the one divergence is qwen3-30b-a3b-2507 on chat-metered-responsiveness, where the generator reports +8% off and this rule finds no dominator within 10% of its speed.