Meta Muse Wonder ships closed-weight, cost near triple per task

Meta launched two frontier-class models on Monday and released neither set of weights: Meta Superintelligence Labs' Muse Wonder and Muse Wonder Full are closed-weight, the flagship is limited to Muse Code subscribers, and the standard model's per-task cost has nearly tripled against its predecessor, to $4.52 per task.

Key facts

• Muse Wonder Full scores 56 on the Artificial Analysis Intelligence Index v4.3 (Meta's own announcement links it), up from 48 for Muse Spark 1.3; Muse Wonder scores 52.
• Muse Wonder pricing is per Artificial Analysis: $4.52 per task versus $1.60 for Muse Spark 1.3, with token prices of $4/M input and $20/M output versus $1.25/$4.25 before.
• Meta trained both models at its gigawatt-scale Prometheus data center.
• Terminal-Bench 4.0 jumps from 33 to 57 percent per The Decoder.
• Both models take text, images, video, and audio input with a 1M-token context window.
• No public weights: VentureBeat confirms neither model ships for download.

What happened

Meta Superintelligence Labs introduced Muse Wonder and Muse Wonder Full on October 12 as its most capable models to date, built on larger base models than the Muse Spark line and trained at the gigawatt-scale Prometheus data center. Both models handle text, image, video, and audio input with a 1M-token context window. Two product hooks come with the release. Active Memory carries context across conversations and recalls it when a task needs it. Active Focus keeps the model working through blockers, trying alternative approaches, and asking for clarification instead of guessing.

Availability splits the family. Muse Wonder is served through the Meta Model API and Muse Code, with Muse app access to follow for subscribers first. Muse Wonder Full exists only behind Muse Code, Meta's paid coding assistant. Nothing is downloadable: VentureBeat notes neither model ships with publicly downloadable weights, a first for a Meta "frontier" label and a clean break from the Llama-era distribution.

The benchmark case is the strongest part of Meta's pitch. Meta reports Muse Wonder Full at 56 on the Artificial Analysis Intelligence Index v4.3, with Muse Wonder at 52, up from 48 for Muse Spark 1.3. The gains concentrate in agentic work: Terminal-Bench 4.0 jumps from 33 to 57 percent, and on the model card's SWE-Marathon agentic coding evaluation, Muse Wonder Full scores 57.5 against 46.3 for Claude Opus 5.5 and 42.5 for GPT-6 Astra. On DeepSWE, the model card reports 79 percent. The claim package comes straight from Meta's own evaluations supplemented by vendor-supplied and third-party results, so treat the ranking as a claim pending independent confirmation, not a settled fact.

Then there is the bill. Per Artificial Analysis, Muse Wonder runs $4.52 per task on the Intelligence Index against $1.60 for Muse Spark 1.3, a near-tripling of per-task cost. Token pricing rose from $1.25/$4.25 per million input/output tokens to $4/$20. The model also burns 28 percent more input tokens per task, which account for 76 percent of total cost. Against the closed frontier tier, that still lands below Claude Fable 5.1 ($7.63) and Claude Opus 5.5 ($5.98) per task and above GPT-6 Astra ($3.26), Gemini 4 Argon ($1.99), and GPT-6.1 Sol ($0.72). No token pricing exists yet for Muse Wonder Full, since it has no public API.

Why it matters

The release matters in two ways, and each half hits a different decision.

The cost half lands on builders running agent workloads. Agentic coding is where Muse Wonder's gains concentrate, and it is also where per-task cost compounds: a loop that used to cost $1.60 per task now prices near $4.52 before any retry budget. For any deployment sizing agent compute, the 2.8x per-task move is the number that matters, not the leaderboard position. Our inference cost curve coverage explains why this happens mechanically: at the frontier, per-task cost rises faster than per-token price falls, because agents spend more tokens to earn each point.

The weights half lands on the open-weights ecosystem. Meta's frontier flagship is now closed-weight, and the open-weight release cadence for the rest of October belongs to other labs: Mistral AI has promised Mistral Large 4 weights by end of October, with the Hugging Face upcoming-release page showing an Oct 31 ETA while secondary coverage floats Oct 27, a date Mistral has not confirmed. Our Muse Spark 1.3-era coverage and the open-versus-closed frontier comparison both track the same widening spread between what closed frontier tiers charge and what open-weight or decentralized capacity can economically serve. A frontier-class Meta model that no self-hoster, marketplace, or fine-tune shop can ever touch changes the shape of that spread for at least a quarter.

One weaker spot is worth naming. Factual knowledge lags the rest of the package: Muse Wonder Full scores 30 on the AA-Omniscience knowledge test against 46 for Claude Opus 5.5, 43 for GPT-6 Astra, and 42 for Gemini 4 Argon. That gap suggests a smaller base model than the coding scores imply, and it matters for retrieval-heavy deployments. Low hallucination rates partly offset it; per The Decoder, Muse Wonder Full's hallucination rate drops from 33 to 23 percent, second-lowest among models scoring at least 45 index points, with only Gemini 4 Argon (15 percent) lower.

Background

Meta's last two releases went the other way on weights. Muse Spark 1.3, the predecessor, shipped as a lightweight, low-latency tier, and the Muse Spark-era coverage of the open-weight wave documented how much of the competitive pressure on closed API providers came from self-hostable alternatives. The Wonder release reverses that pattern at exactly the moment Mistral AI, StepFun, and Xiaomi are pushing open-weight frontier-class models. VentureBeat's writeup notes the Wonder family was built under Alexandr Wang's Meta Superintelligence Labs organization, with the model card including safety evaluations across cyber, bio-chem, and autonomous-action risks.

The benchmark package itself is a mix. Meta's own model card states that its comparison table combines internally run evaluations, vendor-supplied scores, and third-party results, so the headline rankings rest partly on the vendor's own numbers. Independent aggregator LLM Stats currently ranks Muse Wonder Full near the top of its coding and programming category, which is consistent with the agentic-coding-first framing but is still compiled from the same model card data. The debut on the Vals AI Finance Agent Index, at fifth place and beating all OpenAI models per Vals AI's own post, is the one genuinely third-party result in the set.

What's next

Watch three things. First, whether independent evaluators confirm the agentic gains; Artificial Analysis runs its own Intelligence Index and will publish its own ranking positions rather than Meta's. Second, token pricing for Muse Wonder Full, which should surface if Meta opens it beyond Muse Code; it is still unpublished. Third, the open-weights side of the month: Mistral Large 4's weights window runs to Oct 31, and if those weights land Apache 2.0 like Mistral Large 3, the frontier spread between closed-only flagships and self-hostable frontier-class models widens further. Our agent-API guide tracks which tier actually serves agent loops economically, and this release shifts that calculus.