Data Sources

The public face of Xentir's internal licensing register. A source is used only after its terms are read and recorded here — an unread or unverifiable licence is treated as a no-go, not a maybe.

In use

Epoch AI Benchmarking Hub — epoch.ai/benchmarks. Licence: CC BY 4.0, which explicitly permits redistribution and public display with attribution. Required citation, shown on every page that carries Epoch data: “Epoch AI, 'AI Benchmarking Hub'. Published online at epoch.ai. Retrieved from 'https://epoch.ai/benchmarks' [online resource].” · epoch.ai/benchmarks · Licensed CC BY 4.0 Xentir reads only the benchmarks Epoch itself runs (GPQA Diamond, MATH Level 5, SWE-bench Verified, OTIS Mock AIME, FrontierMath, FrontierMath Tier 4, SimpleQA Verified, Chess Puzzles) plus Epoch's own composite (ECI), downloaded as the zip Epoch publishes — never scraped, never read via a third-party client. Epoch also redistributes dozens of third-party leaderboards in the same bundle (*_external.csv); none of those are read, because Epoch's own licence does not extend to a third party's terms. Each would need its own review first.

In use — vendor pricing

Vendor-published pricing (OpenAI, Anthropic, Google) — official price pages only, no aggregators, re-checked weekly: developers.openai.com/api/docs/pricing, platform.claude.com/docs/en/about-claude/pricing, ai.google.dev/gemini-api/docs/pricing. Every stored price carries its source URL and the date it was read, and is shown as "price as published by {vendor}, checked {date}"; a row not refreshed in over a week is marked stale rather than trusted silently. Google's page previously redirected to an authentication prompt when fetched programmatically and shipped blank for that reason; a later re-check found it reachable again and it is now ingested like the other two — if it starts redirecting again, its prices are left as-is until the next successful weekly read and then marked stale. Several individual models on all three vendors' pages publish a context-length-tiered or per-modality price in a single cell rather than one flat per-token figure; those specific rows are left blank rather than guessing which figure applies, even though the vendor itself is in use. A price Xentir does not have is always left blank, never estimated.

Evaluated and not used

LiveBench — permissively licensed (Apache-2.0 / MIT), but its published datasets have not been updated with current models and its site exposes no live data feed. Kept on file rather than used, in case a live feed reappears.

Artificial Analysis — attribution is free to use their API, but redistribution — publishing their numbers on a public page — is sold as a separate product. Internal research only, unless that is purchased.

LLM Stats — terms could not be independently re-verified; the earlier, more conservative reading (personal, non-commercial viewing only) stands.

Hugging Face leaderboard datasets — the platform's terms are not the same as any individual dataset's terms, and none of the underlying datasets have been separately cleared.

Review

Sources are re-verified on a schedule, and a stale review is treated as a no-go until re-read. Questions about a source or a licence: [email protected].