Rank and filter the ~700 tracked model variants — the default starting point.

Answers "which model is smartest / cheapest / fastest / best value right now". Returns a compact markdown table sorted by sort (default: intelligence).

Args: mode: "rank" (default), "value" (intelligence per $/task), or "frontier" (the Pareto intelligence-vs-$/task frontier — the same code path as aa_pareto, which also plots other x axes; the footnote gives the knee). sort: metric/alias to sort by — e.g. intelligence, cost_per_task, tps, ttft, context_tokens, price_in, price_out, params_b, release_date, or any benchmark key. Use aa_reference topic=fields for all names. fields: comma list of columns to show instead of the default set. top: rows to return (default 10; >15 drops the Creator column for brevity). creator: substring match on the creator/lab (e.g. "anthropic"). name: substring match on the model name/slug. open_weights: True/False to keep only open- or closed-weights models. reasoning: True/False to keep only reasoning or non-reasoning models. min_intelligence: drop models below this Intelligence Index value. max_cost_per_task: drop models above this $/task (cache-aware). min_speed: drop models below this median output tok/s. max_ttft: drop models above this median time-to-first-token (seconds). min_context: drop models with a smaller context window (tokens). released_after: ISO date (YYYY-MM-DD); keep models released on/after it. include_deprecated: include deprecated rows (default False). refresh: bypass the cache and re-fetch (use for "released today").

One model's full digest: specs, indices, benchmark scores, pricing and speed.

Use after aa_leaderboard/aa_search when you need everything about a single model. The header and footnote always carry the round-trippable slug.

Args: query: model name or slug (fuzzy). Ambiguous names return a candidate table. sections: comma string (or list) of identity/indexes/benchmarks/pricing/ speed/cost/openness/modalities, or all. Default: identity, indexes, benchmarks, pricing, speed. Alias: specs = identity. Unknown names error with the valid list. refresh: bypass the cache and re-fetch.

The Pareto frontier: any score vs any cost/size/speed axis, with the knee.

Always available — no per-benchmark tool needed. Returns the models where no other model has both a lower-or-equal x and a higher-or-equal score, ordered by x ascending, plus a knee line and coverage.

The three site views this covers:

• aa_pareto (default) = "Intelligence Index vs. Cost per Index Task"

• aa_pareto x=out_tokens_per_task = "Intelligence Index vs. Output Tokens per Task"

• aa_pareto x=index_total_cost = "Intelligence Index vs. Cost to Run the AA Intelligence Index" (whole suite)

Args: score: any rankable with per-model numbers (default intelligence); e.g. cyberIndex, terminalBench40, gpqa, capabilities.legal, openness. x: the trade-off axis — cost_per_task (default), index_total_cost, out_tokens_per_task, time_per_task, context_tokens, params_b, ttft, price_in, price_out. top: frontier points to return (default 15; cheapest first). creator: substring match on the creator/lab. open_weights: True/False to keep only open- or closed-weights models. min_intelligence: drop models below this Intelligence Index value. min_context: drop models with a smaller context window (tokens). include_deprecated: include deprecated rows (default False). include_all: return every frontier point instead of the top cheapest. refresh: bypass the cache and re-fetch.

Side-by-side comparison of 2-12 models on any benchmark (site parity).

Mirrors artificialanalysis.ai's comparison pages: models as columns (or a per-metric delta view), every rankable metric, and a reproducible YAML spec. Rows AA publishes only on an eval chart set / homepage panel (the cyber boards, AA-3D, …) are joined in and the source is footnoted; any cell that stays - is footnoted (a per-model missing on requested metrics: … count, the per-metric detail while the gap is small, or a whole-model "no published pricing or speed" line) so no gap is silent.

Media/speech arena models (e.g. "Grok Voice 3.0") also resolve: comparing only arena entries renders their board metrics (Elo / STS / WER, appearances, price, release, open weights) with a footer warning that per-board scores are not cross-comparable; mixing a language model with an arena entry is refused with an explanation.

Args: queries: 2-12 model names or slugs (language models, or arena entries on one board). metrics: "all" (every rankable key any selected model has), "benchmarks" (benchmark/index kinds only), or a comma list of keys/labels/aliases; empty uses the default rows (Intelligence, Intel/$, $/task, in/out $/1M, tok/s, TTFT, Context). For an arena comparison the names are arena fields (elo, sts, wer, ci, appearances, price, release, open_weights, rank, board, creator). view: "table" (models as columns) or "diff" (one row per metric, value plus the delta vs the first model, e.g. 57.6 (+4.9)). compact: default metric set only — keep the core rows instead of the full set; ignored (and stated in the footer) whenever metrics= is given. spec: a YAML or JSON document with models/metrics/view/compact/filters, merged with the explicit arguments; every reply ends with a copy-pasteable spec: block that reproduces the call (view and compact included, so the round-trip is lossless). refresh: bypass the cache and re-fetch.

Search everything rankable: benchmarks, indexes, derived metrics, arenas.

Each hit carries its How to query hint (the exact tool call), so finding "Artificial Analysis Speech to Speech Index" immediately gives you aa_media board=speech_to_speech.

Args: query: tokenised AND-match over key/label/aliases/description/tags (e.g. "speech", "cyber", "long context"). domain: one domain name (coding, math, science, security, agentic, tool-use, long-context, instruction-following, knowledge, reasoning, safety, multimodal, speech, media, music, video, business, finance, legal, medical, engineering, economics, cost, speed, openness, ...). tag: one raw tag (a narrower filter than domain=). verbose: also show what each thing measures.

Rank models on one benchmark, stating the source and coverage honestly.

Auto-selects the richest source: the full 700-row leaderboard when the field exists there, else AA's curated eval chart set, else the homepage panel; the header says which and warns when it is not a full leaderboard.

Args: benchmark: benchmark key, label or eval slug (e.g. gpqa, gpqa-diamond, "GPQA Diamond", briefcaseElo). top: rows to return (default 15). fields: comma list of columns to show instead of the default set. creator: substring match on the creator/lab. open_weights: True/False to keep only open- or closed-weights models. min_intelligence: drop models below this Intelligence Index value. include_estimates: include AA-estimated scores (default True); estimates are marked with * and footnoted. refresh: bypass the cache and re-fetch.

Provider endpoints: per-host pricing, precision, speed and feature flags.

Pass model for one model's serving endpoints (with per-endpoint $/task and precision), or host/filters for the 1,200-row provider board. With neither, returns the cheapest endpoints globally.

Args: model: model name or slug — switches to per-model endpoint detail. host: substring match on the serving host (e.g. "groq"). top: rows to return (default 15). sort: cost_per_task (default), price_in, price_out, tps, ttft, intelligence or context_tokens. Lower-is-better keys sort ascending. fields: comma list of columns; supports precision, cache_read, discount, e2e, function_calling, json_mode, context_tokens. open_weights: keep only endpoints for open- or closed-weights models. min_context: drop endpoints with a smaller context window (tokens). refresh: bypass the cache and re-fetch.

Human-preference arenas: image, video, speech, music (Elo + 95% CI).

Args: board: any image/video/TTS/music board key (text_to_image, image_editing, text_to_video, image_to_video, video_editing, text_to_speech, text_to_speech_controlled, music_instrumental, music_with_vocals) or a speech board (speech_to_text, speech_to_text_streaming, speech_to_speech — also accepted as realtime_voice / "realtime voice"). Hyphens/spaces/case are ignored (text-to-image == text_to_image). top: rows to return (default 15). fields: comma string (or list) of columns; supports ci_lower, ci_upper, ci (95% CI), open_weights, url. Elo CI is not shown by default. open_weights: keep only open- or closed-weights entries. refresh: bypass the cache and re-fetch.

AA agent leaderboards: coding agents, chatbots and search APIs.

Args: board: coding_agents (default), chatbots or search_api. top: rows to return (default 15). refresh: bypass the cache and re-fetch.

Cheap name -> slug resolver; call this first whenever a name is uncertain.

Always returns the slug so the next tool call can be exact. Media/speech arena models (image/video/speech/music boards) are searched too and returned with their board, so a name like "Grok Voice 3.0" resolves instead of coming back empty.

Args: query: fuzzy model name or partial slug (e.g. "claude opus"). creator: optional creator/lab substring filter. open_weights: True/False to keep only open- or closed-weights models. top: rows to return (default 15). refresh: bypass the cache and re-fetch.

Static reference text: tool routing, metric/eval definitions, provenance.

No network. Use to learn what the fields mean and which tool answers what.

Args: topic: overview (default), indexes, benchmarks, pricing, speed, providers, media, fields (alias list), sources, or all.

Raw JSON escape hatch (token-expensive — prefer the table tools).

Args: kind: model, leaderboard, providers, providers_board, media, agents, home, eval or catalog. query: slug/board/slug argument for the chosen kind (e.g. a model slug for "model", a board key for "media"/"agents", an eval slug for "eval"). fields: comma list of top-level keys to keep per record. top: records to return, capped at 5 (default 1). refresh: bypass the cache and re-fetch.