What are the strengths of this model?

Long-agentic coding plus active memory, tuned for multi-file refactors, delegation and recovery Tends to say it does not know instead of hallucinating, improving honesty on uncertain tasks Frontier-class API at the same $4/$20 per 1M tokens as Claude Opus 5.5

What are the weaknesses of this model?

Artificial Analysis Intelligence Index 52 trails Opus 5.5 (58), GPT-6 Astra (53) and Gemini 4 Argon (53) Omniscience general-knowledge score of 30 is well below Opus 5.5 (46) and Astra (43) Roughly $4.52 per Intelligence Index task — about 3x its own Spark 1.3 ($1.60) — from burning 28% more input tokens

What are the best use cases?

Long-agentic coding and multi-file refactoring Persistent personal assistant built on active memory High-trust document workflows in legal and finance that need a frontier-class API

MetaProprietary

Meta Muse Wonder

Name: Meta Muse Wonder
Author: Meta

Compare this model

Meta's Muse Wonder, released Oct 12, 2026, is Meta's first frontier-class model since Alexandr Wang rebuilt the AI org. It pairs long-agentic coding with active memory and is priced at $4/$20 per 1M tokens — identical to Claude Opus 5.5 — yet scores 52 on the Artificial Analysis Intelligence Index versus Opus 5.5's 58. Its Omniscience general-knowledge score of 30 trails Opus 5.5's 46, so Meta charges frontier prices for a model that scores like a challenger.

Share:X Hatena LINE

Parameters

undisclosed

Context Window

License

Proprietary

Release Date

2026-10-12

Benchmark Performance

AA Intelligence Index

52.0

LMArena Elo

—

HLE

—

ARC-AGI-2

—

SWE-bench Verified

—

GPQA Diamond

—

MMLU-Pro

—

LiveCodeBench

—

AIME 2025

—

MATH-500

—

API Pricing

Input Price (per 1M tokens)

Output Price (per 1M tokens)

$20

Billing Mode: per 1M tokens

Strengths

•Long-agentic coding plus active memory, tuned for multi-file refactors, delegation and recovery
•Tends to say it does not know instead of hallucinating, improving honesty on uncertain tasks
•Frontier-class API at the same $4/$20 per 1M tokens as Claude Opus 5.5

Weaknesses

•Artificial Analysis Intelligence Index 52 trails Opus 5.5 (58), GPT-6 Astra (53) and Gemini 4 Argon (53)
•Omniscience general-knowledge score of 30 is well below Opus 5.5 (46) and Astra (43)
•Roughly $4.52 per Intelligence Index task — about 3x its own Spark 1.3 ($1.60) — from burning 28% more input tokens

Use Cases

•Long-agentic coding and multi-file refactoring
•Persistent personal assistant built on active memory
•High-trust document workflows in legal and finance that need a frontier-class API

Deep Analysis

Artificial Analysis Intelligence Index

Wonder 52 / Full 56

vs Opus 5.5 58, GPT-6 Astra 53, Gemini 4 Argon 53; Wonder ranks #9 of 219 models

Price (Wonder)

$4 / $20 per 1M tokens

identical to Claude Opus 5.5; Wonder Full gated to Muse Code $50/$200 plans, no API

Context Window

1M tokens

multimodal input (text + image); closed weights, not open like Llama

Terminal-Bench 4.0

57% (Wonder Full)

up from 33% on Muse Spark 1.3; ~30% fewer output tokens

SWE-Marathon

57.5 (Wonder Full)

long-run coding; vs Opus 5.5 46.3; vendor-run at xhigh effort

AA-Omniscience (general knowledge)

30 (Full) / 25 (Wonder)

vs Opus 5.5 46, Astra 43, Argon 42; hallucination rate 23% (down from 33%)

Strengths

・Genuine long-agentic coding lead: Wonder Full wins SWE-Marathon 57.5 vs Opus 5.5 46.3 and ties Claude Code on the Coding Agents Index (68).
・Real legal-agent strength (Harvey LAB-AA 13.1 vs Opus 5.5 4.2) and a standout MemoryBench 85.5/85.6 for cross-session recall.
・Active Memory and Active Focus are productized well: it persists key facts and keeps working toward a goal when a step fails, asking before irreversible actions.

Weaknesses

・Priced at Opus 5.5 rates ($4/$20) while scoring six points lower on the independent index (56 vs 58) - you pay frontier money for a challenger score.
・Cost per task is ~$4.52 on the Intelligence Index vs $1.60 for its own Spark 1.3 predecessor (nearly 3x) because it burns 28% more input tokens per task.
・Thin general knowledge (Omniscience 30) and the full safety report is listed as coming soon, not at launch - a model built to remember personal details deserves evidence on day one.

Competitor Comparison

Model	Arena	SWE	GPQA	Price
Claude Opus 5.5	58 (AA Intelligence)	46.3% (SWE-Marathon)	N/A	$4 / $20 per 1M
GPT-6 Astra	53 (AA Intelligence)	42.5% (SWE-Marathon)	N/A	$10 / $50 per 1M
Gemini 4 Argon	53 (AA Intelligence)	N/A	N/A	~$1.99 per task

Overview

Muse Wonder and Muse Wonder Full are Meta's first frontier-class models since Alexandr Wang rebuilt the AI org, launched October 12, 2026. Wonder ships on the Meta Model API, OpenRouter and Muse Code at $4/$20 per million tokens; Wonder Full is held behind Muse Code subscriptions ($50/$200 a month) with no API. Both are multimodal-input, 1M-context, closed-weight models aimed at long agentic coding and personal/professional assistants. On the independent Artificial Analysis Intelligence Index, Wonder scores 52 (rank 9 of 219) and Wonder Full 56, against 58 for Claude Opus 5.5. Meta's own table shows Wonder Full winning 7 of 14 rows, including an 11-point lead over Opus 5.5 on multi-hour coding. The weak spot is factual knowledge: Wonder Full's Omniscience score is 30 versus 46 for Opus 5.5, though its hallucination rate fell to 23%. Zuckerberg told the FT Meta has caught up at the frontier and rejected a slowdown - but the index says it caught up on price, not on the factual bench.

Benchmarks & Performance

Detailed Comparison

Muse Wonder's only honest comparison is against the frontier it is priced to join. At $4/$20 it matches Claude Opus 5.5 exactly while scoring six points lower on the index, and it undercuts GPT-6 Astra ($10/$50) while scoring at the same 53 level as Astra on the composite. | Feature | Muse Wonder | Claude Opus 5.5 | GPT-6 Astra | | :--- | :--- | :--- | :--- | | **Price (in/out)** | $4 / $20 | $4 / $20 | $10 / $50 | | **AA Intelligence** | 52 / 56 (Full) | 58 | 53 | | **Context** | 1M | 1M | 1.05M | | **Omniscience** | 25 / 30 | 46 | 43 | | **Weights** | closed | closed | closed | **The call.** Wonder is the best Meta coding/agent model yet and a credible buy if your workload is long agentic loops. For factual Q&A or research, the Omniscience gap says wait. And the $4/$20 rate buys a 52-index model when Astra scores 53 at $3.26 a task.

Community Feedback

Launch-day reaction split on the price. The Decoder's independent read credited Wonder Full as a top-tier coding and agent model but flagged the Omniscience gap and the delayed safety report. Developer forums noted the $4/$20 rate equals Opus 5.5 while the index is lower, reading it as Meta pricing itself into the frontier club before the benchmarks fully agree. * **What landed.** The SWE-Marathon and Harvey LAB legal-agent wins drew real respect; Active Memory was called a genuinely useful agent feature. * **What drew pushback.** The Omniscience score (30) and the missing day-one safety report; several reviewers said a model designed to remember personal details across months should ship its guardrail evidence at launch. * **The Zuckerberg quote.** His FT claim that Meta has caught up and rejects a slowdown was read against the six-point index gap as confidence leading the data.

Use Cases

**1. Long agentic coding runs** * Example: a multi-hour refactor across many files with sub-agent delegation. * Why Wonder Full: 57.5 on SWE-Marathon and a Coding Agents Index of 68; built for planning, delegation and error recovery. **2. Legal and finance document work** * Example: contract review and evidence analysis with the AutoClaw agent platform. * Why Wonder: Harvey LAB-AA 13.1 leads Opus 5.5 (4.2) by a wide margin on Meta's own legal benchmark. **3. Persistent personal assistants** * Example: an assistant that keeps context across a long project via Active Memory. * Why Wonder: MemoryBench 85.5/85.6 is the highest Meta has measured for cross-session recall. **When to pick something else** * Opus 5.5 or Gemini 4 Argon for factual research and general Q&A (Omniscience gap). * An open-weight model if you need on-prem control - Wonder is closed, unlike Llama.

Latest News

**Launch (October 12, 2026).** Meta released Muse Wonder and Muse Wonder Full, its first frontier-class models since the Wang reorganization. Wonder is available on the Meta Model API, OpenRouter and Muse Code; Wonder Full is limited to Muse Code Power Usage ($50/month) and Maximum Usage ($200/month) plans with no public API. **Pricing.** Wonder lists at $4 per million input and $20 per million output tokens - identical to Claude Opus 5.5. Wonder Full has no token price. **Positioning.** Meta frames both around frontier coding, long-running agents, advanced computer use and Active Memory. Zuckerberg told the Financial Times that Meta has caught up at the frontier and rejected calls for a slowdown. The full safety report is listed as coming soon rather than shipping at launch.

On the independent Artificial Analysis Intelligence Index, Wonder scores 52 (rank 9 of 219) and Wonder Full 56, against 58 for Claude Opus 5.5. Meta's own table shows Wonder Full winning 7 of 14 rows, including an 11-point lead over Opus 5.5 on multi-hour coding. The weak spot is factual knowledge: Wonder Full's Omniscience score is 30 versus 46 for Opus 5.5, though its hallucination rate fell to 23%. Zuckerberg told the FT Meta has caught up at the frontier and rejected a slowdown - but the index says it caught up on price, not on the factual bench.

Sources

Analysis generated: 2026-10-13