Methodology

Assumptions & Methodology

Every number on this site is either measured from data we collect, or assumed with a stated value. This page lists all of them so you can audit, disagree, and re-run with your own inputs.

updated August 05, 2026 · data 2024-11-09 → 2026-08-05 (635 days, 792 models)

1 · Where the data comes from

source	covers	how
OpenRouter frontend stats	last ~10–31 days, exact	`model-activity` per model: prompt/completion tokens, requests, tool calls — full per-model granularity. Scraped daily.
OpenRouter author pages	~90-day rolling, per model	Each lab page embeds a chart of daily tokens for its top models. Parsed for the recent window.
Internet Archive (Wayback)	back to 2024-11-09	Historical snapshots of the same author pages, stitched together to extend history ~1.5 years.
OpenRouter app pages	snapshot (daily fwd)	Per-app top-20 model breakdown — the demand side. Totals, not a daily series.
Model list / pricing	current	Per-model prompt & completion list prices, free flag, creation date.

2 · The assumptions (and their values)

assumption	value	rationale & how it's used
Gross token margin assumed	90%	Inference cost ≈ 10% of list price. Used for all "gross profit" figures. Linear lever — swap freely.
OpenRouter take-rate assumed	5–10%	Credit fees + BYOK fee + provider spread on GMV. We show a 5 / 7.5 / 10% range rather than one number.
OR → total-API scale assumed	×10	OpenRouter is estimated at ~10% of a major lab's API traffic. Used only for "est. total revenue". The biggest single uncertainty — measured "via OR" figures don't use it.
Frontier training cost assumed	$100M	Order-of-magnitude compute cost of one Opus-class run. Excludes R&D, staff, serving. Used for model ROI.
Prompt:completion mix (history) assumed	recent actual	Author-page history gives total tokens only. For historical revenue we apply each model's real recent prompt:completion split (from the exact data) at its actual list price. Recent-window revenue is fully exact.
Useful-life threshold assumed	50% of peak	A model's "useful life" = launch → first day usage falls below half its peak. Choice of threshold; 50% is conventional.
Smoothing method	7-day MA	Centered 7-day moving average on all time-series to remove weekend seasonality.
Partial-day trim method	<40% of median	The current UTC day is incomplete; trailing days below 40% of the weekly median are dropped so charts don't dip to zero.
Free / preview exclusion method	excluded	Free and preview variants spike then vanish (misleading). Excluded from launch/fade/head-to-head analysis.
China vs West classification	by lab origin	DeepSeek/Qwen/MiniMax/Moonshot/Z.ai/Xiaomi/Tencent/StepFun/ByteDance = China; OpenAI/Anthropic/Google/xAI/Meta/Mistral/Microsoft = West. Stealth/unknown (OpenRouter's Owl) excluded.

3 · Known limitations

Historical absolute totals undercount. Author pages expose only each lab's top ~12 models; the long tail ("Others") is dropped. Relative shares and trends are robust; absolute T/day in the deep past is a lower bound.
App data is a snapshot, not yet a time-series — app-blend history accrues from first capture forward (plus a few Wayback points).
OpenRouter ≠ the whole market. It's the largest neutral router, but direct-API and first-party app traffic (ChatGPT, Claude.ai) aren't here. Vendor totals are the OpenRouter slice.
Revenue is list-price GMV, not billed revenue — discounts, committed-use, and BYOK aren't visible.

How to read the site: trust the measured columns as fact; treat assumed figures as a transparent frame you can re-dial. Every "est. total" / "profit" / "OpenRouter revenue" number is downstream of the assumptions above.