Independent · coding agents
WinnerAA Coding Agent Index
Artificial Analysis v1.5: DeepSWE, Terminal-Bench 4.0 and SWE-Atlas-QnA, each model in its vendor’s own agent. Muse Wonder Full in Muse Code ties Claude Sonnet 5.5 in Claude Code for first.
SPIRITT
Muse WonderMuse Wonder. Built to remember, and to keep going.
Meta’s new frontier model is built for personal and professional agents that carry context forward and push through blockers. Independent tests put Muse Wonder Full within two points of the top model, and level with the best coding agents.
Launch day: Meta, the evaluators and early testers.
Meta’s announcement, independent results and first hands-on tests, including mixed and skeptical ones. Community claims are the creators’ own and were not reproduced by SPIRITT.
A bigger base model for agents that keep going.
Independent evaluators confirm a genuine frontier model. Muse Wonder Full leads several agentic and reasoning tests, while Claude Sonnet 5.5 and Gemini 4 Argon, both missing from Meta’s scorecard, lead others.
What Meta launched on October 12, 2026
Meta introduced two models. Muse Wonder (muse-wonder-1.0) is in preview on the Meta Model API and available in Muse Code. Muse Wonder Full, the higher-capability variant, is only in Muse Code on the Power Usage and Maximum Usage plans, and Meta has not published a per-token price for it. Meta says Muse Wonder will soon power its Muse app, starting with subscribers.
Meta built the release around Active Memory, which carries important context across conversations and recalls it when it becomes relevant, and Active Focus, which keeps working toward a hard goal: trying alternatives, asking for clarification and checking in before important actions. Meta’s internal tests report 85.6% on MemoryBench for Muse Wonder and 54.2% on FocusBench for Muse Wonder Full.
On the API, Muse Wonder takes text, image, video, audio and PDF input, returns text, has a 1,048,576-token context window and supports reasoning effort from minimal to max. Standard pricing is $4 per million input tokens, $0.20 cached and $20 output, with no long-context premium: about 3.2x the input and 4.7x the output price of Muse Spark 1.3.
Artificial Analysis scores Muse Wonder Full at 55.8 on its Intelligence Index, third among distinct models, and Muse Wonder at 52.4, up from 48.1 for Muse Spark 1.3. Muse Wonder Full also leads SWE-Marathon, GDP.pdf and CritPt. It trails on Terminal-Bench 4.0, AutomationBench-AA and Vals AI’s finance benchmark, and Artificial Analysis flags knowledge recall as a relative weakness.
This page is a release brief. Muse Wonder is not in the SPIRITT model picker at publication; Muse Spark 1.3 is available in SPIRITT today.
Released Oct 12, 2026Muse Wonder + Wonder Full1M-token context$4 / $20 per 1M tokens55.8 AA Intelligence IndexRelease brief
Artificial Analysis Intelligence Index
Near frontierArtificial Analysis v4.3.2, October 12, 2026, max reasoning except Gemini 4 Argon (high). Muse Wonder Full is third among distinct models, 0.2 points behind Claude Sonnet 5.5, which Meta’s launch chart leaves out. Both Muse Wonder models gain on Muse Spark 1.3.
Where Muse Wonder leads, ties and trails
Independent results come first, with the full field of competitors. Meta-reported measures follow and stay labeled; some cite third-party sources that had not published Muse Wonder results at retrieval.
Independent · marathon coding
WinnerSWE-Marathon v1.1
Abundant AI public leaderboard: substantial projects, eight binary-graded trials each, each charted model in its own agent at max reasoning (Muse Spark 1.3 at xhigh). Meta and Abundant’s launch post report 57.5% for Full at xhigh, a run the public board did not list at retrieval.
Independent · PDF-grounded work
WinnerGDP.pdf
Artificial Analysis all-pass score, each model at its best setting on the Intelligence Index page. A narrow lead of 0.4 points. Meta’s table cites Surge, whose public board had no Muse rows at retrieval.
Independent · research-level physics
WinnerCritPt
Artificial Analysis. Muse Wonder Full leads a tight cluster by 1.1 points; Muse Wonder trails it.
Independent · real-world knowledge work
Near frontierGDPval-AA v2.1 Elo
Artificial Analysis Elo on economically valuable tasks. Muse Wonder Full is 9 points behind Claude Opus 5.5 and 176 points above Muse Spark 1.3.
Independent · long-horizon coding
Near frontierDeepSWE v1.1
Artificial Analysis average pass@1, each model in its vendor’s agent; Gemini 4 Argon is not publicly available. Meta’s table cites 79.0% and 78.8% from Datacurve, whose public board had no Muse rows at retrieval.
Independent · business workflows
AutomationBench-AA
Artificial Analysis strict pass rate, the metric Meta cites. Meta’s chart shows Muse Wonder Full first because it omits Gemini 4 Argon and Claude Sonnet 5.5. On the partial-credit score it ranks third at 71%.
Independent · terminal coding
Terminal-Bench 4.0
Artificial Analysis accuracy. Muse Wonder Full trails Claude Sonnet 5.5 by about 6.5 points but gains 24 points on Muse Spark 1.3.
Independent · GDP-weighted agent work
Vals Index v2.1
Vals AI finance, coding, legal and tax index. Muse Wonder Full ranks fifth, at $8.56 per test, the lowest cost among the top five.
Independent · financial analysis
Vals Finance Agent v2
Vals AI. Muse Wonder Full ranks fifth. Meta’s chart shows it first because its comparison set excludes Gemini models and earlier Muse Spark versions.
Independent · cost to run
Cost per AA Index task
Artificial Analysis, lower is better. Meta’s token prices plus high input use put Muse Wonder mid-pack. Muse Wonder Full has no public token price, so no cost is listed.
Meta-reported · memory
WinnerMemoryBench
Meta’s internal test of recall across 64K to 100K-token conversation histories with distractors and updated facts. Comparison set chosen by Meta; not independently reproduced.
Meta-reported · persistence
WinnerFocusBench
Meta’s internal test of multi-step personal tasks in areas like housing, travel, IT and finance, with real-world blockers. Muse Wonder trails GPT-6 Astra and Claude Opus 5.5.
Meta-reported · personal work composite
WinnerPersonal Superintelligence Index
Meta’s internal composite of seven evaluations, including FocusBench, JobBench and AutomationBench-AA. Comparison set chosen by Meta.
Meta-reported · legal agents
Harvey LAB-AA v1.1
Muse rows are Meta-reported, citing Artificial Analysis, whose public page did not list Muse Wonder at retrieval. Other rows are from that public page. A task counts only if it meets every criterion with no hallucination.
Meta-reported · professional tasks
Near frontierJobBench
65 tasks across 35 occupations, as reported by Meta. Muse Wonder Full is 0.1 points behind Claude Opus 5.5.
Meta-reported · computer use
OSWorld V2.1
Partial-credit score as reported by Meta; Claude Opus 5.5’s figure is Anthropic’s provider-reported score. On strict completion, Muse Wonder Full scores 39.4% against 43.1% for Claude Fable 5.1.
Meta-reported · health questions
HealthBench Pro
525 physician-rubric health questions, answered without tools. Claude scores come from Anthropic’s system cards.
Meta-reported · from-scratch programs
ProgramBench
Share of rollouts passing at least 95% of hidden tests, as reported by Meta, with every model in mini-swe-agent. The widest gap on Meta’s own table.
Meta-reported · privacy
WinnerMishandling personal information
Violation rate in Meta’s safety testing, lower is better. Claude Opus 5.5 is not shown in Meta’s chart.
Meta-reported · adversarial robustness
WinnerExecuting unauthorized code
Attack success rate in Meta’s safety testing, lower is better. A narrow edge over Claude Fable 5.1; Claude Opus 5.5 is not shown.
Independent sources, retrieved October 12, 2026: Artificial Analysis Intelligence Index v4.3.2, its component evaluations and Coding Agent Index v1.5; Abundant AI’s SWE-Marathon v1.1 leaderboard; Vals AI Finance Agent v2 and Vals Index v2.1. Charts show selected current models, not full leaderboards. Artificial Analysis charts use each model’s best setting listed on its Intelligence Index page, max reasoning unless labeled; Claude results there use Artificial Analysis’s max-with-fallback setting. Meta-reported charts come from Meta’s launch post and model card, where the comparison set is Meta’s choice and competitor runs are best effort; MemoryBench, FocusBench and the Personal Superintelligence Index are internal. Leaderboards change. API prices, evaluation costs and SPIRITT plan prices are different things.
Start the long-running work today.
Bring the goal into SPIRITT, build with Muse Spark 1.3 now, and compare new Meta models on the same work as access changes.
01
Bring the real goal
A move to plan, a claim to file, a codebase to ship. Give SPIRITT the files, context and success criteria, in a workspace with a browser, terminal, memory and integrations.
02
Build with Muse Spark 1.3 today
Muse Spark 1.3 is in the SPIRITT model picker now. Meta built Muse Wonder and Muse Spark to accept the same API requests, so the workflows you build today are ready for a side-by-side comparison as access changes.
03
Keep the context. Keep going.
Long-running agent work needs durable context. A SPIRITT Workspace keeps your files, memory and sessions in one place, so agents can pick up where they left off. Test any model against your longest, messiest task.
Start the long task today.
Open a SPIRITT Workspace and build with Muse Spark 1.3 and the other models in your picker. Bring the goal, the files and the context; the workspace keeps the work moving.
Questions
Muse Wonder FAQ
01What is Muse Wonder?+
Meta’s frontier model family for personal and professional agents, released October 12, 2026. It includes Muse Wonder and the higher-capability Muse Wonder Full, and focuses on long-running agent work, coding, memory across conversations and persistence through blockers.
02What is the difference between Muse Wonder and Muse Wonder Full?+
Muse Wonder (muse-wonder-1.0) is in preview on the Meta Model API and in Muse Code. Muse Wonder Full is only in Muse Code on the Power Usage and Maximum Usage plans and has no published per-token price. Setting reasoning effort to max on Muse Wonder does not select Full. On the Artificial Analysis Intelligence Index, Full scores 55.8 and Muse Wonder 52.4.
03How much does Muse Wonder cost?+
On the Meta Model API: $4 per million input tokens, $0.20 per million cached input tokens and $20 per million output tokens, with no long-context premium. Web search adds $2.50 per 1,000 queries. Artificial Analysis measured $4.52 per Intelligence Index task, lower than Claude Fable 5.1, Opus 5.5 and Sonnet 5.5, but higher than GPT-6 Astra, Gemini 4 Argon, GPT-6.1 Sol and Muse Spark 1.3. These are Meta’s prices, not SPIRITT pricing.
04Does Muse Wonder beat Claude, GPT and Gemini?+
It depends on the test. Independently, Muse Wonder Full ties Claude Sonnet 5.5 for first on the Artificial Analysis Coding Agent Index, leads SWE-Marathon, GDP.pdf and CritPt, and ranks third among distinct models on the Intelligence Index, behind Claude Opus 5.5 and Claude Sonnet 5.5. It trails on Terminal-Bench 4.0, AutomationBench-AA, the Vals Index and the Vals Finance Agent benchmark, where Claude or Gemini models lead.
05Are Meta’s benchmark results independent?+
Partly. Meta mixes third-party results, provider-reported scores and its own runs. Its 14-test scorecard compares only Claude Fable 5.1, GPT-6 Astra and Claude Opus 5.5. At retrieval, several cited sources had not published matching Muse Wonder results: Artificial Analysis’s Harvey LAB-AA page, Datacurve’s DeepSWE board and Surge’s GDP.pdf board. On SWE-Marathon, Meta and evaluator Abundant report 57.5% for Full at xhigh, while Abundant’s public board listed 53.1% for Full and 52.5% for Muse Wonder at max.
06What are Active Memory and Active Focus?+
Meta’s names for the two behaviors it trained for. Active Memory carries important context about you across conversations and retrieves it when relevant. Active Focus keeps working toward a hard goal, trying alternatives, asking for clarification when needed and checking in before important actions.
07How large is the context window?+
1,048,576 tokens on the Meta Model API. Muse Wonder accepts text, image, video, audio and PDF input and returns text. Reasoning effort runs from minimal to max; turning reasoning off is not supported.
08How does Muse Wonder compare with Muse Spark 1.3?+
Meta positions Muse Wonder for long, complex work where capability matters more than cost, and Muse Spark for high-volume or cost-sensitive tasks. Artificial Analysis scores Muse Wonder 52.4 against 48.1 for Muse Spark 1.3, with a large jump on Terminal-Bench 4.0, at about 2.8x the cost per index task.
09What are the known limitations?+
Meta’s model card lists errors in multi-step reasoning on novel tasks, weaker visual and spatial reasoning, especially 3D, untested languages, and results that vary with serving setup, prompts, tools and agent harness. Artificial Analysis flags knowledge recall as a relative weakness, offset by low hallucination rates. Muse Wonder is in preview, and availability varies by country.
10How does Meta address safety for agents?+
Meta calls these its strongest safeguards to date against chemical, biological, radiological, nuclear and cyber misuse. In Meta’s tests, Muse Wonder Full mishandled personal information in 2.2% of cases and Muse Wonder had the lowest unauthorized-code attack success rate shown. Without mitigations, Meta does not rule out high risk in chemical and biological domains; with mitigations it assesses moderate or lower risk. Meta recommends human confirmation before irreversible agent actions.
11Can I use Muse Wonder in SPIRITT?+
Not yet. This page is a release brief, not an availability announcement: Muse Wonder is not in the SPIRITT model picker at publication. Muse Spark 1.3 is available today, so you can build Meta-model agent workflows in a SPIRITT Workspace now.
Buy from builders who use what they sellBuilt using