Meta's Muse Wonder launch splits frontier access between API and subscription

Two models, two doors

Two models shipped on 12 October. They do not travel the same road. Muse Wonder runs on the Meta Model API and inside Muse Code, Meta's terminal coding agent [3][6][12].

Muse Wonder Full is the one behind glass. It is limited to Muse Code subscribers and has no per-token listing at all [3][6][12]. The API rates for the accessible model are $4 per million input tokens and $20 per million output tokens, with cached input billed at $0.20 per million [3][11][12]. Muse Code costs $50 a month on the Power plan or $200 on Maximum [1][11]. The Muse app's own consumer subscriptions sit lower, at $20 or $100 a month [1][10]. Both models take a one-million-token context window and accept text, images, video and audio as input, with text out [2][11].

Meta's pitch for the pair is memory and persistence. Active Memory carries user context across conversations, and Active Focus keeps the model working on hard tasks, trying alternatives and checking in before important actions [2][8][12]. Neither feature is sold apart from the model that carries it.

Why a subscription wall matters more than a benchmark

Benchmarks sell a launch. Access rules decide who gets to build on it.

Because Muse Wonder Full has no per-token price, API-only teams can reach just the smaller model [6][12]. That inverts the usual pattern. The flagship is normally the one you rent by the token. A team that wants to test Meta's best model on its own workload has to buy a coding subscription to get at it [6][12]. Meta's previous release, Muse Spark 1.3, was pitched as competitive rather than frontier; Muse Wonder is the first framed as frontier class [3][6]. The company spent fourteen months rebuilding its AI operation to get here, with Alexandr Wang hired to run Meta Superintelligence Labs [3][10][12].

The commercial backdrop is a company spending hard. Meta stock rose about 4% in afternoon trading on launch day [4].

Meta reported Q2 revenue of $60.8 billion, up 28% [4]. It expects 2026 capital expenditure of $130 billion to $145 billion, against free cash flow of $784 million in the quarter [4]. Demand looks real: the Muse app reached 5 million downloads in 22 days, faster than ChatGPT, Grok and Claude, according to Sensor Tower estimates cited by Yahoo Finance [10]. Safety is part of the sales pitch too. Meta says its strongest safeguards against cyber misuse ship with the paid model [1].

The scoreboard, and where it leaks

Muse Wonder Full scores 56 on the Artificial Analysis Intelligence Index v4.3 and Muse Wonder 52, against 48 for Muse Spark 1.3 [2][7]. Claude Opus 5.5 leads at 58, with GPT-6 Astra and Gemini 4 Argon at 53 [7]. Agentic work is where the jump shows. On Terminal-Bench 4.0, performance rose from 33% to 57% [2][7]. Muse Wonder Full scores 68 on the Coding Agents Index, tying Claude Code with Sonnet 5.5, while Muse Wonder scores 63, up from 54 [2]. On ProofBench, Muse Wonder Full solved 94 of 100 problems, up from 58 [2].

Factual knowledge is the weak spot. Muse Wonder Full gets 30 on AA-Omniscience and Muse Wonder 25, against 46 for Claude Opus 5.5, 43 for GPT-6 Astra and 42 for Gemini 4 Argon [2]. Hallucination fell from 33% to 23% for Muse Wonder Full, the second-lowest rate among models scoring at least 45 index points [2].

On Vals AI, Muse Wonder Full debuts fifth of 47 models [2][6]. Muse Wonder debuts tenth [2][6].

Meta's own table is kinder in places and harsher in others. MemoryBench puts Muse Wonder at 85.6, top of the five models listed, while ProgramBench shows 50.0 against 70.5 for Claude Opus 5.5 [11]. Several of those rows are internal Meta evaluations, and the model card says competitor runs are best-effort and may not be tuned [11]. The benchmark results were supplied by Meta and have not been independently verified [1].

The price of climbing

Muse Spark 1.3 launched at $0.10 per million input tokens and $0.20 per million output [6]. Muse Wonder charges $4 and $20, roughly a forty-fold jump on input [6].

Per task, Muse Wonder costs $4.52 on the Artificial Analysis Intelligence Index, against $1.60 for Muse Spark 1.3, $3.26 for GPT-6 Astra and $1.99 for Gemini 4 Argon [7]. Alexandr Wang told CNBC the model is still priced meaningfully below the leading OpenAI and Anthropic models [7]. On the per-task numbers, that holds against Anthropic. It does not hold against OpenAI or Google.

Wang's framing is broader than coding. "These are true frontier models," he told Bloomberg. "We really think they are competitive with the best models out there, and particularly state of the art on personal and professional agentic work" [10].

The launch was not clean. It shipped with bugs, including one that miscounted usage limits for Muse Code subscribers, according to the FT [7].

What happens next

Muse Wonder will power the Muse app itself [2][8][12]. Paying subscribers get it first [2][8][12].

Meta says it plans to build specialized models from Muse Wonder for all Muse users over time [2][12]. Muse Wonder Full stays off the API for now [6][12]. Its cost per task remains unpublished. Meta has not said when, or whether, it will reach the API. Both models were trained at Meta's Prometheus data center in central Ohio [3][6].