Muse Wonder and Muse Wonder Full landed on October 12, 2026. They are Meta Superintelligence Labs' first frontier model release under Alexandr Wang, according to TechCrunch (Russell Brandom) and VentureBeat (Carl Franzen).

I am Nitin Kumar, Lead Project Coordinator at Debugged Pro. India and USA agency product teams keep asking the same thing: can we put Wonder on the Meta Model API today, or is Full stuck behind Muse Code? This guide sticks to verified press and Meta facts. No invented quotes. No fake bench numbers.

Muse Wonder Meta frontier model graphic for India and USA agency teams — Feature graphic: Muse Wonder and Muse Wonder Full, Meta's first frontier models. Original Debugged Pro illustration.

What Muse Wonder and Muse Wonder Full Actually Are

TechCrunch says Muse Spark aimed to be a lightweight, low-latency model. The Wonder family is larger and more capable, with stronger results on complex coding. Meta's model card, as summarized by TechCrunch, puts Wonder and Wonder Full in a range comparable to Fable 5.1 and GPT-6 Astra on Meta's own testing.

Meta's research blog states: "Together, they mark the next step on our scaling ladder and a major leap toward personal superintelligence." Both models also power the Muse agent. Training happened at Meta's Prometheus data center in central Ohio, a gigawatt-scale site that came online earlier this year.

This is the first frontier release since Wang restructured Meta AI in August of last year. Prior MSL releases included Muse Spark and the Muse agent. Meta stayed out of the state-of-the-art model race OpenAI and Anthropic usually lead, until this week.

Muse Wonder Access Paths: API vs Muse Code Full

Standard Muse Wonder ships on the Meta Model API and inside Muse Code. Model id: muse-wonder-1.0. Interfaces are OpenAI and Anthropic compatible, which cuts migration cost for existing agent stacks.

Muse Wonder Full is Muse Code subscribers only. It is not on the API. VentureBeat lists Muse Code Full tiers as Power Usage at $50/mo (20x Everyday) and Maximum Usage at $200/mo (80x Everyday). Agencies that need Full inside a customer-facing product cannot host it on Meta Model API today.

Weights are proprietary. Unlike the earlier Llama open era, Wonder parameters are not downloadable for self-hosting. That is a hard stop for buyers who require on-prem weights.

Muse Wonder verified pricing and access facts for agencies — Muse Wonder facts: API price, cache, web search, rate limits, Wonder Full Muse Code tiers. Original Debugged Pro illustration.

Muse Wonder API Pricing and Rate Limits

Wonder API list pricing is $4 per million input and $20 per million output tokens. Cached input is $0.20 per million. Web search adds $2.50 per 1,000 queries. TechCrunch notes the token rate sits roughly like GPT-6 Sol or Opus 5.5; VentureBeat places Wonder beside Opus 5.5 on list rates and below GPT-6 Astra and Fable 5.1.

Documented rate limits: 3,000 requests per minute and 4 million tokens per minute. Context sits near one million tokens. Multimodal input covers text, image, audio, and video. Output is text.

Long-running agents burn tokens fast. Artificial Analysis estimates about $4.52 to run its average Intelligence Index task on Wonder, versus $5.98 for Opus 5.5 and $3.26 for GPT-6 Astra. Cheap list rates do not always mean cheap completed jobs.

Benchmarks: Artificial Analysis and Meta Model Card

Artificial Analysis Intelligence Index (independent): Wonder Full 56 (ties Claude Sonnet 5.5), ahead of GPT-6 Astra 53, Gemini 4 Argon 53, and Claude Fable 5.1 53, behind Claude Opus 5.5 at 58. Standard Wonder scores 52, one point behind Astra.

Meta's model card (max reasoning), as reported by VentureBeat, includes these rows. Attribute scores to Meta's card plus Artificial Analysis where noted. Higher is better within each row. Do not compare scores across different benches as if they share a scale.

Benchmark	Wonder Full	Wonder	GPT-6 Astra	Claude Opus 5.5
Artificial Analysis Intelligence Index	56	52	53	58
Harvey LAB-AA (legal)	13.1	10.0	8.6	4.2
AutomationBench-AA	45.4	39.1	41.6	41.2
JobBench	75.3	70.5	66.1	75.4
DeepSWE 1.1	79.0	78.8	74.1	74.2
SWE-Marathon	57.5	53.1	42.5	46.3
Terminal-Bench 4.0-AA	57.1	52.0	59.1	59.6
HealthBench Pro	62.3	60.5	65.3	65.6

Wonder Full looks strong on sustained coding (SWE-Marathon) and some professional agent tasks. Opus still leads the AA aggregate. Opus and Astra beat Wonder Full on Terminal-Bench 4.0. Health scores favor Opus and Astra. Meta mixes internal evals, vendor scores, and third-party results, so run your own golden tasks before you promise clients.

Active Memory, Active Focus, and Domain Claims

Meta pushes Active Memory: keep what matters across days, weeks, and months. Research scientist Kavya Kopparapu said: "We put a tremendous amount of effort into making sure your model can remember what's important to you over days, weeks, months, and beyond." On MemoryBench, Meta testing puts Muse Wonder above Fable, Opus, or Astra.

Active Focus aims to keep agents on goals when plans fail. Agents can revise plans, try alternatives, and ask permission before consequential actions. Meta also claims domain strength in law, finance, and health, plus frontier coding. Multimodal input and long context support those agent loops.

For agencies, memory is a product choice, not a slogan. Decide what the agent may store, how clients erase it, and what breaks under zero data retention.

Data Modes, Retention, and Safety Notes

Meta Model API Wonder runs in Standard mode: submitted content is not used for training. Muse Code defaults to Contributor mode, which can train on prompts, code, files, context, and outputs. Subscribers can disable Contributor mode; the change applies to future submissions only. Meta's terms bar confidential or personal data while Contributor is on.

Zero data retention is available to qualified orgs, with trade-offs. VentureBeat notes ZDR setups may drop server-side conversation state, background requests, file uploads, or built-in web search. That collides with Active Memory demos. Confirm feature availability before you sell a memory-heavy agent.

Meta's official Muse Wonder post cites CBRN and cyber misuse safeguards and its strongest alignment claim to date, with a full safety report coming soon. The model card also covers safety evals. Meta still recommends human confirmation for consequential agent actions. Models can still fail on unfamiliar multi-step work.

What India and USA Agencies Should Do Next

Most clients will not care about Meta's scaling ladder. They care about API access, cost, data use, and whether Full is even reachable. Use a short, checkable brief.

Split API Wonder from Muse Code Full. Full is not on Meta Model API today.
Price real agent loops. Include $4/$20 tokens, $0.20 cache, and $2.50 per 1k search queries.
Lock Standard mode for production API work. Audit Muse Code Contributor defaults before client code lands there.
Design Active Memory with erase paths. Map ZDR trade-offs if legal requires them.
Require human confirm for consequential actions. Meta itself recommends it.
Benchmark on your golden suite. AA and Meta card scores are guides, not contracts.
Assign owners. Product, security, and delivery should share one diligence sheet.

Agency checklist for Muse Wonder API and Muse Code Full rollout — Agency checklist after Muse Wonder launch: access path, pricing, data mode, Active Memory, human confirm, NIST. Original Debugged Pro illustration.

When the talk turns into a delivery redesign, point clients to software development consulting with a clear agent integration workstream. A model launch alone is not an architecture.

For related agent ops coverage, see our AI agent monitors agency guide and the Cloudflare Deno acquisition agency guide. More delivery notes live on the Debugged Pro blog.

Map Muse Wonder Diligence to NIST AI RMF

Regulated buyers in India and the USA often want a framework they already know. Tie Muse Wonder client questions to the NIST AI Risk Management Framework:

Govern: who approves Meta Model API spend, who owns Contributor vs Standard mode, who can pause Muse agents.
Map: which prompts, code, and memory stores touch Wonder, and which data classes cannot leave the client tenant.
Measure: golden-task pass rates, token cost per completed job, MemoryBench-style recall checks, and incident history.
Manage: human confirmation gates, dual-path failover to another frontier API, contract exit rights, and ZDR feature trade-offs.

This keeps security and product from talking past each other. It also stops the false story that a Meta frontier score removes delivery risk.

Client Conversation Script You Can Reuse

Keep the first update short. Boards want checkable facts.

What happened: Meta released Muse Wonder and Muse Wonder Full on October 12, 2026, as MSL's first frontier models under Alexandr Wang (TechCrunch, VentureBeat, Meta blog).
Why it matters: Wonder is on Meta Model API at $4/$20 per M with Active Memory claims; Full scores higher on AA (56) but stays Muse Code only.
What is limited: no open weights; Full not on API; Contributor mode defaults in Muse Code; ZDR may drop memory-related features.
What we will change here: update client briefs on API vs Full, data modes, agent human-confirm gates, and NIST AI RMF diligence for India and USA buyers.

Key Takeaways

Muse Wonder is Meta's first frontier release. Wonder + Wonder Full; Oct 12, 2026; Meta Superintelligence Labs / Alexandr Wang.
API path is Wonder only. muse-wonder-1.0 at $4/$20 per M; cached $0.20; search $2.50/1k; Full is Muse Code Power/Maximum only.
Scores sit near the frontier. AA: Full 56, Wonder 52; Meta card shows strong SWE-Marathon and legal agent rows with gaps elsewhere.
Active Memory is the product pitch. Kopparapu quote on recall over days to months; MemoryBench claim vs Fable, Opus, Astra (Meta testing).
Proprietary, not open weights. Hosted access only; Standard API mode vs Muse Code Contributor defaults matter for client code.
Agencies should update diligence. Access path, cost, data mode, human confirm, and NIST controls for India and USA clients.

FAQ: Muse Wonder and Agency Diligence

What is Muse Wonder and how is it different from Muse Spark?

Muse Wonder and Muse Wonder Full are Meta Superintelligence Labs' first frontier models, released October 12, 2026. TechCrunch reports they are larger and more capable than Muse Spark, with better results on complex coding. Both power the Muse agent.

Is Muse Wonder Full available on the Meta Model API?

No. Wonder Full is limited to Muse Code subscribers and is not served through the API, per TechCrunch and VentureBeat. Standard Muse Wonder is available via the Meta Model API (model id muse-wonder-1.0) and Muse Code.

How much does Muse Wonder cost on the API?

Wonder costs $4 per million input tokens and $20 per million output tokens. Cached input is $0.20 per million. Web search is $2.50 per 1,000 queries. Rate limits are 3,000 requests per minute and 4 million tokens per minute, according to VentureBeat citing Meta.

Are Muse Wonder weights open like earlier Llama models?

No. VentureBeat notes both Wonder models are proprietary and not released with publicly downloadable weights, unlike Meta's earlier open Llama approach. Access is through Meta's hosted API or Muse Code.

What should India and USA agencies check before piloting Muse Wonder?

Confirm API vs Muse Code Full access, Standard vs Contributor data modes, Active Memory retention rules, human confirmation for consequential agent actions, and map controls to the NIST AI Risk Management Framework Govern, Map, Measure, and Manage functions.

About the author: Nitin Kumar is Lead Project Coordinator at Debugged Pro. He works with India and USA agency teams on delivery plans for AI products, agent integrations, and client-ready risk briefings.