Muse Wonder Full: A new Fable level model on artificial analysis

The model card has the numbers and the gaps. I'll stick to those figures, then check section length and the deck before sending the article.Meta released Muse Wonder Full (Max) in October 2026 as a proprietary reasoning model. artificialanalysis.ai's report scores it 56 on the Artificial Analysis Intelligence Index, 4th of 699 models. The card lists text, image, speech, and video input, text output, and a 1M token context window. Speed and token prices are unpublished because the page says there is no publicly available API.

Developers comparing models can read this as the full card: the October 2026 release, the score versus the class median, the ten tests in Intelligence Index v4.3.2, the 1M window, and the blank API, speed, and price. The card publishes no API endpoint and no dollar rate.

What shipped in Muse Wonder Full

Muse Wonder Full (Max) is Meta's proprietary model, released in October 2026. This page shows the reasoning version. Reasoning is set to Yes, and a non-reasoning variant may also exist. Inputs are text, image, speech, and video. Output is text. Context is 1M tokens, about 1500 A4 pages in size 12 Arial, in a class of 699 models. Meters show 4 out of 4 units for intelligence, 3 out of 4 for verbosity, and unknown out of 4 for speed and cost.

The Intelligence Index score is 56, rank 4 of 699, above the median of 12. Artificial Analysis says Muse Wonder Full (Max) "is amongst the leading models in intelligence and well priced when comparing to other models of similar price." The run generated 140M output tokens, verbosity rank 142 of 699, against a median of 88M, called somewhat verbose. Speed is N/A, with tokens per second unknown. Cost is N/A: input and output are dashes, and cost per Intelligence Index task is unknown.

What improved in Muse Wonder Full

No earlier Muse score, cache-hit price, or cache-write price is printed. Against the class median, the index is 56 versus 12, and output tokens are 140M versus 88M. Intelligence rank is 4 of 699 (4 out of 4 units) and verbosity rank is 142 of 699 (3 out of 4 units). Input, output, speed, and cost per task are blank, and the page says there is no publicly available API. Higher is better for the index and for tokens per second. Lower is better for USD cost per task.

Intelligence Index v4.3.2 uses ten evaluations: AA-Briefcase v1.1, GDPval-AA v2.1, AutomationBench-AA, Terminal-Bench 4.0, SciCode, Humanity's Last Exam, GDP.pdf, CritPt, AA-Omniscience, and AA-LCR v1.1. Updated AA-Briefcase v1.1 scores agentic knowledge work as (Elo-500)/2000. Elo mixes rubric pass rate, analytical quality Elo, and presentation Elo through synthetic head-to-head matches. Elo and the 95% confidence interval are clamped at 0. Updated GDPval-AA v2.1 scores agentic real-world work on that scale. SciCode (coding) and CritPt (physics reasoning) are under review. Separate Muse scores on the ten are not shown.

What you gain from Muse Wonder Full

The card pairs rank 4 of 699 with text, image, speech, and video input, text output, and a 1M token window (about 1500 A4 pages at size 12 Arial). Capability indexes are Finance & Accounting, Strategy & Ops, Legal, Engineering, and Economics. Finance & Accounting uses AA-Omniscience, GDPval-AA v2.1, AA-Briefcase v1.1, Humanity's Last Exam, AutomationBench-AA, AA-LCR v1.1, and GDP.pdf, charted for 27 of 210 models, with no public API. Intelligence generally translates across use cases, while specific evaluations may be more relevant for certain use cases.

AA-Omniscience runs from -100 to 100. Correct answers are rewarded, hallucinations are penalized, and a refusal draws no penalty. Zero means as many correct answers as incorrect ones. A negative score means more incorrect than correct. The non-hallucination rate is 1 minus the hallucination rate. Charts cover 27 of 573 models for AA-Omniscience and 27 of 228 models for AA-Briefcase. Tokens per second and USD per task are unknown, so the score of 56 and the 140M output tokens have no speed or bill attached.

How to get Muse Wonder Full

Controls on the page are Compare, Try it out, API Provider Benchmarks, and "Add model from specific provider." Intelligence, speed, and cost per task each say "No publicly available API." No base URL, model id, install command, session switch, or pinned launch flag is printed. Input and output are dashes, so the blended 3:1 input/output price ratio has no Muse rate. Proprietary models share a price band with open-weights models: under $0.15, $0.15 to $1, or over $1 per 1M tokens. Muse Wonder Full (Max) is in none of those bands.

Open-weights classes are Tiny at or below 4B parameters, Small from 4B to 40B, Medium from 40B to 150B, and Large above 150B. Muse Wonder Full is proprietary, and no parameter count is given, so those classes do not place it. Commercial Use Restricted means commercial use is limited. Non-commercial means the license prohibits commercial use. Neither label is applied. Reasoning models are compared with both reasoning and non-reasoning models. Non-reasoning models are compared only with other non-reasoning models. No update command or default model is listed.

What to watch after Muse Wonder Full

A non-reasoning variant may also exist. Speed is N/A, tokens per second are unknown, and cost per Intelligence Index task is unknown. Where a price exists, cost adds input, cache-hit, cache-write, reasoning, and answer token prices, divides by task count, and weights by Intelligence Index weight. Muse has no point on the cost, time, or output-token charts. Charts of 27 of 699 models draw a Pareto line and a most attractive quadrant beside StepFun, Z AI, Xiaomi, MiniMax, OpenAI, Anthropic, Mistral, Meta, SpaceXAI, NVIDIA, Google, Alibaba, DeepSeek, and Kimi.

Nineteen of 28 intelligence evaluations are listed. Also named: Harvey LAB-AA v1.1 (updated, legal agentic work, hallucination-gated all-pass rate), EnterpriseOps-Gym-AA (agentic business operations), Terminal-Bench-Science 0.1 (new, scientific research in a terminal), FrontierCode 1.1 (new, software engineering in open-source repositories), AA-AnalystAgent (spreadsheets and documents), ITBench-AA (Kubernetes root-cause analysis), MMMU-Pro (visual reasoning), and MLCR-AA (medical long context). AutomationBench-AA is agentic SaaS workflows, Terminal-Bench 4.0 is agentic coding and terminal use, GDP.pdf is all-pass document reasoning, AA-LCR v1.1 is long context reasoning, and Humanity's Last Exam is reasoning and knowledge. Labels include Coding, Agentic, Tool Use, Private Dataset, User Interaction, Finance, Medical, Legal, Intelligence Index, Long Context, Multimodal, Instruction Following, Faithfulness, Writing, and Business.