Artificial Analysis Top Lists

Extract ranked AI model and coding-agent records from public Artificial Analysis leaderboards. Eight independently ranked lists cover intelligence, output speed, task cost, intelligence-to-cost value, balanced intelligence/speed, Coding Agent Index, coding-agent value, and coding-agent completion time.

Model-only runs fetch one model page. Coding-agent runs fetch the coding-agent page only when at least one coding-agent list is requested. Mixed runs fetch each page once and give all records one run timestamp.

Lists

ID	Metric	Order	Eligibility
intelligence	Intelligence Index	Highest first	Finite intelligence
speed	Median output tokens/second	Highest first	Finite speed
costPerTask	Intelligence Index cost/task, USD	Lowest first	Finite cost, including zero
intelligenceCostRatio	Intelligence Index / cost/task	Highest first	Finite intelligence and cost > 0
intelligenceSpeed	Geometric mean of intelligence and speed percentile ranks	Highest first	Finite intelligence and speed > 0
codingAgentIndex	Coding Agent Index × 100	Highest first	Finite index
codingAgentValue	Coding Agent Index points / mean cost/task	Highest first	Finite index and cost > 0
codingAgentSpeed	Mean agent wall time/task, seconds	Lowest first	Finite time >= 0

intelligenceSpeed ranks eligible models independently on intelligence and output speed. Each rank is converted to 0–100 using (number below + half of tied peers) / (eligible count - 1) × 100; tied values receive the same midpoint rank. A one-model population receives 100. Final score is sqrt(intelligencePercentile × speedPercentile). Calculations and sorting use unrounded source values.

Primary metric ties use deterministic display-name/identity ordering. maxItems applies after filtering and sorting, independently per list.

Input

{
  "lists": ["intelligenceCostRatio", "intelligenceSpeed", "codingAgentIndex", "codingAgentValue", "codingAgentSpeed"],
  "maxItems": 10,
  "creatorFilter": "Anthropic",
  "modelFilter": "Claude",
  "agentFilter": "Claude Code",
  "includeDeprecated": false
}

Field	Type	Default	Meaning
lists	array	intelligence, speed, costPerTask	One to eight unique list IDs. Input order controls output blocks.
maxItems	integer	10	Records per list, 1–100.
creatorFilter	string	omitted	Creator substring; applies to both source types.
modelFilter	string	omitted	Model name/slug substring; applies only to model lists.
agentFilter	string	omitted	Agent name/display label/provider substring; applies only to coding-agent lists.
includeDeprecated	boolean	false	Include deprecated models; irrelevant to coding-agent lists.

Filters are trimmed, NFKC-normalized, case-insensitive literal substrings. Relevant filters compose with AND semantics. Empty filters are omitted.

Output

Every dataset item is flat. Existing intelligence, speed, and costPerTask model records retain their prior fields and metric behavior. New model-list records use the same model shape with new metricName, metricValue, and metricUnit metadata.

Coding-agent example:

{
  "list": "codingAgentValue",
  "rank": 1,
  "metricName": "codingAgentValue",
  "metricValue": 60,
  "metricUnit": "indexPointsPerUSD",
  "sortDirection": "descending",
  "sourceType": "codingAgent",
  "agentId": "agent-run-id",
  "agentName": "Claude Code",
  "agentDisplayName": "Claude Code - Sonnet",
  "provider": "anthropic",
  "hostModelSlug": "anthropic_claude-sonnet",
  "hostModelName": "Sonnet",
  "agentCreator": "Anthropic",
  "modelCreator": "Anthropic",
  "codingAgentIndex": 60,
  "costPerTaskUsd": 1,
  "timePerTaskSeconds": 100,
  "benchmarkCount": 3,
  "benchmarkNames": "DeepSWE, SWE-Atlas-QnA, Terminal-Bench v2",
  "benchmarkRowsJson": "[{...}]",
  "sourceUrl": "https://artificialanalysis.ai/agents/coding-agents",
  "fetchedAt": "2026-07-22T13:49:59.964Z"
}

benchmarkRowsJson contains the complete normalized benchmark breakdown as a JSON string, keeping dataset columns flat and CSV-safe. Parser resolves inline and referenced Next Flight rows, including forward/path references split across chunks. It rejects highlights-only data, unresolved/cyclic/conflicting references, duplicate identities, incomplete benchmark sets, inconsistent weighted aggregates, and payloads below sanity floors.

Default key-value store record OUTPUT reports parsed, filtered, generated, and pushed counts. Each pushed dataset record uses existing model-result Pay-Per-Event event. Charge-limit handling stops before counting rejected record.

Data behavior

• Direct public SSR requests; no browser, login, private API, proxy, or pagination.
• Semantic model table remains fallback for current models where selected fields exist.
• Deprecated requests and metrics unavailable in fallback fail instead of returning partial data.
• Valid filters with no matches return empty output; unfiltered metric loss fails as schema drift.
• Results represent current source snapshots, not historical data.
• Independent Actor; not affiliated with or endorsed by Artificial Analysis.