DC Power is our weighted aggregate of raw benchmark scores from LiveBench, Epoch AI, ARC Prize and Terminal-Bench. It is not an ECI rescaling. Power = 0.35 × Coding + 0.35 × Reasoning + 0.30 × Agents. These are explicit editorial preferences, not statistically learned weights or a probability of success.
Source balance. At full coverage, LiveBench contributes 46.5%, original Epoch AI runs 28%, ARC Prize 10.5% and Terminal-Bench 15%. One source does not gain weight by publishing more rows. ECI, vendor reports and external scores republished through Epoch have zero Power weight. Direct ARC and Terminal results enter once.
Epoch reasoning panel. Equal weights for original Epoch-run Chess Puzzles, FrontierMath tiers 1–3 v2 private and OTIS Mock AIME 2024–2025. All three are required. GPQA and other supplementary tests stay in the matrix; vendor-only GPQA reports do not become independent Epoch runs.
Missing evidence stays visible. Power, Eco and Overall require every fixed source panel. Domain scores can show a partial summary: available panel weights are renormalized, with coverage and missing inputs shown. A partial score is not a win against a complete score. The same complete panel is required for every ranked model.
Fixed editions and effort policy. LiveBench 2 edition 25 October 2026 and Terminal-Bench 4.0. Select known effort in this order: max, xhigh, high, medium, low, minimal, none, declared highest-supported. Never select the largest numerical result. A dated later run can replace an earlier run under the same setup; unresolved conflicts or different harnesses are missing.
Publisher independence is limited. Four separate source organizations do not imply blind or independently reproduced runs, independent funding, equal compute budgets or uncorrelated tests. Terminal-Bench measures a model with its published agent; Claude Code, Codex and other harnesses differ. Inspect the selected configurations. Public benchmark data cannot establish outcomes in your application.
No invented confidence interval. We do not convert ECI uncertainty into uncertainty for this aggregate. No statistically valid DC Power interval is available. Raw means are averaged on their original 0–100 scales; test difficulty and contamination can affect results. A small aggregate difference is descriptive, not proof of superiority.
Token cost C = (3,000 × input rate + 1,000 × output rate) / 1,000,000, uncached. DC Eco = Power / (1 + 100 × C), eligible only for complete Power ≥ 50 and known prices. At $0.01, Eco is half Power. DC Overall = 0.75 × Power + 0.25 × Eco. Use the comparison calculator for your workload and quality weight.
Cost is conditional. Prices are listed USD per million tokens. Reasoning tokens, tools, retries, cache writes, discounts and long-context surcharges are excluded. These are token-spend estimates, not measured cost per successful task. Quality benchmark effort does not establish a matching token workload.
v3 replaces the previous ECI transform; scores across methodology versions are not directly comparable. Data refreshes every six hours, not on each page view. Benchmark editions are pinned for comparability; changing the test panel requires a methodology revision. Imports older than 24 hours are marked delayed. Catalogue releases continue to update independently. Azure listings do not confirm account quota or regional capacity.
Epoch data: CC BY 4.0. Other publishers: attributed public leaderboard score facts. DARC.DEV defines the aggregate and makes no claim of source endorsement. No private account details are published.