Meta's first frontier model, Muse Wonder, is proprietary and priced to undercut OpenAI and Anthropic

FACT: On 12 October 2026, Meta announced the Muse Wonder series, two proprietary models — Muse Wonder and Muse Wonder Full — described as its most capable models to date, built for "personal and professional agentic tasks" and positioned on Meta's "scaling ladder" toward "personal superintelligence". The announcement was published simultaneously across the AI at Meta research blog, the Meta Newsroom, and the developer platform (dev.meta.ai).

FACT (availability and commercial terms):

• Muse Wonder is available immediately through the Meta Model API and Muse Code at $4 per million input tokens, $0.20 cached input, $20 per million output tokens (dev.meta.ai). It is OpenAI-SDK-compatible (change base_url, model id muse-wonder-1.0).
• Muse Wonder Full is not served through the API; it is available only inside Muse Code, on the Power Usage plan and a new Maximum Usage plan offering 80x the Everyday Usage limit (dev.meta.ai; TechCrunch).
• CNBC reports consumer access to the model in the Muse personal-agent app requires a monthly subscription of $20 or $100, and Muse Code plans run $50 or $200 per month; Meta quotes Alexandr Wang as saying the API tier is priced "meaningfully below" leading OpenAI and Anthropic models because it is "a much larger model and at frontier level".

FACT (model basics, per the Muse Wonder model card): input modalities text, image, video, audio; output text; 1M-token context window at inference; trained on multilingual data; released 12 October 2026 by Meta Platforms, Inc.

FACT (lineage): the models build on work laid down by Muse Spark, introduced 8 April 2026 as the first step of the scaling ladder; Muse Spark 1.3 (released about six weeks before Wonder) was the previous top of the Muse lineup, running the Muse personal-agent app since its US launch on 8 September 2026.

COMPANY CLAIM: Meta states Muse Wonder and Muse Wonder Full "deliver state-of-the-art performance on personal and professional agentic tasks" and that they were trained at its "gigawatt-scale Prometheus data center". Meta further claims (per its model card, relayed by TechCrunch) that both models achieve benchmark scores comparable to Claude Fable 5.1 and GPT-6 Astra; some benchmark claims in the developer post (e.g., Muse Wonder Full "tops" DeepSWE and SWE-Marathon in "independent third-party runs") have not been reproduced by the third parties named.

INTERPRETATION: This is not "a new generative model direction" in the narrow sense of a research prototype — it is a commercially released proprietary frontier model family, the first of its kind from Meta, aimed squarely at the agent market rather than at chat. The discovery framing understated the event.

• Meta now fields a frontier model. Before Muse Wonder, Meta's most capable public release was the lightweight Muse Spark line, which Meta itself positioned as a low-latency small model, not a state-of-the-art entry. TechCrunch frames Wonder as "the company's first attempt at a frontier model" since chief AI officer Alexandr Wang restructured Meta's AI efforts in August 2025. Independent measurement (Artificial Analysis Intelligence Index v4.3) puts Muse Wonder Full at 56 (max effort) and Muse Wonder at 52, versus Muse Spark 1.3's 48 — a material jump in a six-week span.
• Meta's flagship is now proprietary, not open-weight. Muse Wonder and Muse Wonder Full are closed, API/subscription-only models; neither can currently be self-hosted (VentureBeat). This diverges from the Llama open-weight tradition and puts Meta's frontier line in direct competition with OpenAI, Anthropic and Google on their own commercial terms.
• The product surface changed: the Muse personal-agent app (launched 8 September 2026, free tier) will be powered by Muse Wonder "soon", with subscribers getting it first; Muse Code (Meta's coding agent, launched August 2026) gains the two new models including a new Maximum Usage plan with an 80x usage limit.
• The user-facing economics changed: Meta is charging explicitly for frontier capability — $20/$100 per month in the Muse app, $4/$20 per million tokens on the API — a notable step for a company whose AI has historically been free/ad-funded (CNBC emphasises this as a test of Meta's ability to create AI-derived revenue outside ads).
• What did not change yet: no safety report is published (Meta says a full report "with additional safety evaluations will be available soon"); Muse Wonder is not yet live in the Muse app; and no open-weight version of the family has been announced.

FACT (what Meta describes — the two headline behaviours):

• Active Memory: carrying important context about the user across conversations and actively retrieving the details relevant to the task at hand — the claim is about when earlier information becomes relevant, not merely about long context windows. Meta evaluates this with MemoryBench, an internal benchmark testing recall of critical details across long, complex multi-turn conversations of 64K–100K tokens, with distractor conversations and injected contradictory/updated facts (per the model card; TechCrunch links the MemoryBench paper at arxiv.org/abs/2510.17281 and quotes researcher Kavya Kopparapu: "a tremendous amount of effort into making sure your model can remember what's important to you over days, weeks, months, and beyond").
• Active Focus: sustained effort toward a challenging goal — trying alternative approaches, working through blockers, adapting to new information, asking for clarification when needed, and checking in before taking important actions. Evaluated through FocusBench (internal personal-agentic eval across housing, travel, IT, finance) and SWE-Marathon (ultra-long-horizon software work).

COMPANY CLAIM (training methodology): the series extends the Muse Spark work "across pretraining and reinforcement learning" on the scaled stack, trained at the gigawatt-scale Prometheus data center. Meta describes a generalisation analysis showing that, at equal in-distribution post-training performance, the larger/stronger base models behind Wonder Full generalise better to out-of-distribution tasks ("big model smell") — presented as evidence that post-training gains carry over to unfamiliar tasks, which Meta argues matters for personal agents.

COMPANY CLAIM (evaluation breadth): Meta reports results across law (Harvey's Legal Agent Benchmark LAB), finance (Vals Finance Agent V2), health (HealthBench Pro), computer use (OSWorld V2.1), terminal coding (Terminal-Bench 4.0), long-horizon coding (DeepSWE 1.1, SWE-Marathon), and its internal Personal Superintelligence Index (PSI — composite of Agents' Last Exam, AutomationBench-AA, FocusBench, JobBench, GDP.pdf, VBT, DRACO).

COMPANY CLAIM (safety): Meta states the family has "our strongest safeguards to date" against CBRN and cyber misuse while preserving frontier cyber capability for defensive use, is "our most aligned model to date", was trained to avoid sharing private information without permission and to refuse harmful requests, and was developed under the Meta Superintelligence Scaling Framework (updated 2 October 2026, extending loss-of-control requirements to training/evaluation). A full safety report is promised but not yet published — an important gap for an EARLY RESEARCH story.

INDEPENDENT EVIDENCE (Artificial Analysis, same-day, 12 October 2026):

• Intelligence Index v4.3 at max effort: Wonder Full 56, Wonder 52, versus Muse Spark 1.3's 48; Wonder Full improves across all ten evaluations. Largest gains are agentic: Terminal-Bench 4.0 33%→57% (level with Gemini 4 Argon high), GDPval-AA v2.1 1681→1857 Elo (#2), AA-Briefcase 1581→1729 Elo (#3). Wonder Full leads CritPt (33%) and GDP.pdf (33%), and ranks #3 on AutomationBench-AA (71%) and SciCode (63%).
• Coding Agents Index at max effort: Wonder Full 68 (level with Claude Code + Sonnet 5.5 max, ahead of Opus 5.5 max 66); Wonder 63; Muse Code with Spark 1.3 scored 54.
• Efficiency: both models use ~54k output tokens per Intelligence Index task (~10% fewer than Spark 1.3); Wonder Full uses 16% fewer input tokens than Wonder. Cost per task for Wonder (max): $4.52 — lower than Claude Fable 5.1 ($7.63), Opus 5.5 ($5.98), Sonnet 5.5 ($5.46), but higher than GPT-6 Astra ($3.26), Gemini 4 Argon ($1.99), GPT-6.1 Sol ($0.72) and Spark 1.3 ($1.60).
• Known weakness: knowledge (AA-Omniscience): Full 30 / Wonder 25, level with Spark 1.3 and well below Opus 5.5 (46), GPT-6 Astra (43), Gemini 4 Argon (42). This is offset by low hallucinations: Full 23%, Wonder 27% — the second- and third-lowest rates among models scoring ≥45 on the index.

INTERPRETATION: Muse Wonder is the single clearest signal yet that Meta intends to compete at the frontier of agentic AI, not just at the lightweight tier. Three things make it matter:

1. It changes the frontier field. A third major US lab now ships a measured frontier agent model with aggressive pricing and a 1M context window. Enterprises that want frontier agents on a hyperscaler-friendly vendor now have a Meta option priced below OpenAI/Anthropic flagship equivalents.
2. It makes agent reliability a product category. "Active Memory" and "Active Focus" reframe the agent problem from raw benchmark score to persistence: remembering across conversations and pushing through real-world blockers. For the agent economy, that is the real operational bottleneck, and Meta has chosen to bet its flagship on it.
3. It is a structural business test for Meta. Meta is asking consumers and developers to pay for frontier AI ($20/$100 app tiers, $4/$20 API, subscription-only "Full"). WSJ frames the release as a test of whether Muse Wonder can improve retention and accelerate adoption of the Muse agent, whose downloads have slowed since launch. CNBC frames it as Meta's attempt to create AI revenue outside ads.

• Verified frontier agent performance on Meta's stack: independently measured top-tier results on agentic knowledge work (GDP.pdf #1), terminal coding (Terminal-Bench 4.0 level with Gemini 4 Argon), and coding agents (Coding Agents Index 68, tied with Sonnet 5.5 in Claude Code).
• Long-horizon agent workflows with a different approval posture: practitioners (e.g., Tech Fabric's analysis) note that a model trained to check in before consequential actions and ask for clarification changes where human approval gates sit — from "every write-back" toward "writes the model flagged, plus a sample". That is where operating savings in agentic automation live, if the behaviour holds in production.
• Cross-conversation memory as a model feature: Active Memory addresses the recurring cost of rebuilding agent state on every run; context that persists and is retrieved when relevant is a model-side answer to a problem teams currently solve with plumbing.
• A consumer-scale test of paid frontier AI: Muse app subscribers get Wonder-powered agents first — the first mass-market test of Meta's willingness-to-pay thesis.
• Defensive cyber capability preserved: Meta claims Wonder keeps frontier cyber capability available for defensive use (finding flaws in your own code) while hardening against misuse — relevant to security teams.

1. When will the promised safety report arrive, and what do its CBRN/cyber evals show at the thresholds of the (2 Oct 2026) Meta Superintelligence Scaling Framework?
2. Will Active Memory and Active Focus behaviours hold in production deployments — i.e., does the check-in-before-action behaviour survive jailbreaks, tool errors and adversarial prompts (the "evaluation awareness" class of issues seen with Muse Spark per Apollo Research)?
3. What are the real retention economics in the Muse app: does Wonder lift the slowed download/retention curve (WSJ's test), and what is the paid-tier conversion rate?
4. When does Muse Wonder actually roll out to Muse subscribers ("soon"), and which specialised models will Meta derive from it next?
5. Will Meta publish effective-context measurements, retention/data-processing terms for the API, and per-service audio support confirmation?
6. How will OpenAI/Anthropic respond on price for agent workloads (Fable 5.1 at $7.63/task now sits under direct pressure)?
7. Is an open-weight version of the Wonder family planned, or is the closed frontier a permanent structural choice?

PREDICTION (near-term, 1–4 weeks):

• Wonder rolls out to Muse app subscribers ("soon") and Meta publishes the promised full safety report; both are company commitments, and both are falsifiable; the safety report date is the marker to watch.
• Advertising and measurement of real-world Active Memory/Focus behaviour in production will appear from practitioners; expect a spike in third-party evaluations once the API is load-tested by the community (AA-style cost-per-task, effective context, and jailbreak/refusal reports).
• A price-response from OpenAI/Anthropic for agent-market pricing is likely given AA cost-per-task pressure ($4.52 vs $7.63/$5.98).
• Further Muse-ladder steps: Meta says Wonder will be used to develop "specialised models" for all Muse users — the next rung of the scaling ladder ("with much more to come").

PREDICTION (medium-term, 1–3 quarters):

• Retention data from the Muse app will determine whether the paid-frontier-AI bet works; WSJ's framing (downloads slowed) is the benchmark for success.
• Expect independent red-teaming of Wonder's cyber/CBRN safeguards to be the next major EV-newscycle item if the safety report's thresholds and the September permission dispute resurface in the press.
• In the enterprise, expect Meta to push Wonder through partnerships and possibly cloud marketplaces to widen distribution beyond first-party.

Meta stopped being a spectator at the frontier. Muse Wonder is a proprietary, commercially priced, independently measured frontier agent model — Meta's first — and it is aimed squarely at the part of the market that is hardest and most valuable: agents that remember, persist and check in. The responsible reading of this EARLY RESEARCH story is carefully calibrated: the release itself is confirmed fact; the "state-of-the-art" framing is company claim; the independent evidence (index 52/56, coding-agents 63/68, low hallucination, weak knowledge, high input-token cost) is real but partial — one evaluator, one day one, no safety report yet. The genuinely new thing is not just the model, but the business experiment wrapped around it: Meta asking millions of consumers and developers to pay for personal-superintelligence agents, backed by the world's largest social-media distribution and the containment machinery of a publicly updated safety framework. Whether the model or the business case is the bigger story will be decided by two numbers no benchmark can supply — Muse app retention and the safety report's publication date.