Methodology

Data

OpenRouter supplies listed endpoint prices, effective prices and model usage. Artificial Analysis supplies benchmark scores and workload token counts.

Usage is prompt plus completion tokens; cached input and reasoning are already included. Through 22 August 2026 we use OpenRouter’s leading-model rankings, including “other.” From 23 August we use finalized daily model activity. The current UTC day is excluded from daily usage charts.

Benchmarks and editions

The AA views use the scores, workloads and weights of the selected Intelligence Index edition. v4.3.2 retains the v4.3 benchmark set, with updated AA-Briefcase v1.1 scoring. Earlier editions remain separate archived series; we do not substitute one edition’s workloads into another.

Individual task views use the named exam’s accuracy or success rate and measured tokens per attempt. Scores run from 0 to 100; changes are reported in points. Omniscience uses its accuracy component.

Historical charts apply retained benchmark measurements to historical prices and usage. They are retrospective, not records of the benchmark results known on each date. Different editions can produce different levels and trends.

ThoughtPrice composites

Current v2 is the default: HLE (20%), CritPt (15%), Omniscience accuracy (20%), AA-LCR (15%), SciCode (15%) and Terminal-Bench v4.0 (15%). The same fixed weights determine both score and token workload.

v1 is retained as a separate historical series. It uses the legacy seven-exam mix below; it is not revised or backfilled with v2 measurements.

The mix allocates 40% to reasoning, 30% to knowledge and documents, and 30% to agentic work. CritPt receives less weight because it is specialized in research physics. These are fixed editorial weights, not estimates of economic value.

All seven score/workload pairs must be present in one captured configuration. Each model’s representative configuration is initially the median-score eligible setting, then frozen for v1. Measurements are also frozen; revised definitions require a new version.

This is an experimental ThoughtPrice measure. It borrows AA’s category balancing, but does not reproduce AA’s index or Epoch’s ECI, which estimates capability statistically across benchmarks.

Task cost

We divide each exam’s canonical token totals by its task count, then weight the per-task quantities where a composite is selected. For the retained legacy AA snapshot, quantities were recovered from reported component costs and prices, with a 10% output-token reconciliation tolerance.

Cost is the sum of each token quantity multiplied by its price: ordinary input, cache reads, cache writes, and answer plus reasoning output. AA’s captured cache-hit rate sets the cache quantities.

The listed cost is the cheapest complete offer from a single endpoint. Missing required cache-read prices exclude an endpoint; missing cache-write prices use ordinary input pricing. Historical prices remain active until superseded, including scheduled price windows. The task explorer samples each completed day’s end.

Listed costs are per attempt, not per successful answer or actual user bill. They exclude external tool and compute charges. Effective pricing is separate: its input rate applies once to all prompt tokens, without adding a second cache discount.

The Tradable Qualified Basket additionally requires observed endpoint uptime of at least 95%, median output speed of at least 20 tokens/second, and at least 1% of that model’s trailing provider token volume. Its history starts with our endpoint-performance and provider-share snapshots.

Chart calculations

Estimated task volume is N = observed tokens / tokens per task. Average capability and task cost use the same eligible models and these same weights. They describe the changing mix of models used, rather than a fixed-model price index.

A fixed-score floor is the cheapest retained model configuration meeting the selected score threshold, regardless of usage. On capability versus cost, the green floor instead follows the usage-weighted average score. The price selector switches between listed offers and qualified endpoints; qualified costs use the eligible subset while the performance line retains the listed model mix.

Score-band cost averages use token-volume weights, or equal-model weights when selected. Capability-demand views can instead use observed request counts or Ramp’s enterprise spend shares.

Model-family charts show seven-day averages. Estimated task volume is summed within each family; estimated listed revenue is sum(N × listed task cost). This reprices observed volume under the benchmark workload mix, rather than measuring provider receipts.

Coverage

Task-cost calculations require a matched model, valid score and workload, and complete eligible prices. Missing measurements are excluded, not filled with zero or another benchmark’s values.

Coverage is the share of observed token volume represented by eligible models. Unmatched models and “other” remain in the denominator. It measures coverage within the observed OpenRouter data, not the global inference market.

AA usage views prefer the default evaluated setting, otherwise the median-score match. TP views freeze their initial representative setting. OpenRouter does not identify the reasoning setting of every request.

Newly covered models can expand historical coverage. A changing model mix or coverage can move an average even without a like-for-like price change; compare the fixed-score floor when asking whether a given capability has become cheaper.