Meta Launches Muse Wonder and Muse Wonder Full: MSL's Next-Gen Frontier Models

Meta Superintelligence Labs (MSL) has released its second-generation model family Muse Wonder (codename "Watermelon") and Muse Wonder Full. This is the lab's second family-level release about 14 months after its founding, and only 6 weeks after the previous generation Muse Spark 1.3.

The availability of the two models is deliberately split: Muse Wonder is offered through the Meta Model API and Muse Code, while Muse Wonder Full is only for Muse Code subscribers. Muse Wonder will also power the Muse personal agent product and will be opened first to subscribers.

Opening the "fast version" to the API while locking the "full version" behind a subscription is the most direct product signal of this release: frontier capabilities first serve the company's own subscription entry point, rather than being rolled out to all callers first.

Artificial Analysis Intelligence Index v4.3 provides the most complete third-party reading so far: Muse Wonder Full scores 56 at highest reasoning effort, and Muse Wonder scores 52, both above Muse Spark 1.3's 48. On the same index, Opus 5.5 leads with 58.

In other words, MSL pushed its own score from 48 to 52 and 56 in 6 weeks, but the top spot still belongs to Opus 5.5. This is a position of "near the frontier" rather than "occupying the frontier."

On evaluations closer to agentic execution, the improvement is more pronounced. On Terminal-Bench 4.0, Wonder Full improves from 33% to 57%, and Wonder from 33% to 52%. In Muse Code's Coding Agents Index, Wonder Full scores 68, tied with Claude Code + Sonnet 5.5 and above Opus 5.5's 66.

On Vals AI's private ProofBench, Wonder Full scores 94/100, compared with Spark 1.3's 58. It should be distinguished that Meta's self-reported benchmarks—SWE-Marathon 57.5%, DeepSWE v1.1 79.0%, Harvey LAB 13.1, MemoryBench 85.5—are vendor claims, not independently verified, and can only be treated as source_claim.

The published specs are: 1 million token context; input supports text, images, video, and audio, output is text only; pricing is $4 per million tokens input, $0.20 cached input, and $20 output.

Only half of the cost side can be calculated. Wonder costs $4.52 per indexed task, driven mainly by high input usage. For comparison, under the same measure, Fable 5.1 is $7.63, Opus 5.5 is $5.98, Sonnet 5.5 is $5.46, Astra is $3.26, Argon is $1.99, Sol is $0.72, and Spark 1.3 is $1.60.

But Wonder Full has no published token price, so its cost cannot be calculated. This means it is currently impossible to judge the unit economics of the "full version" under real workloads—and that is precisely what subscribers care about most.

There was a specific failure on launch day: Wonder and Wonder Full abnormally rapidly consumed Muse Code quotas. The issue was fixed the same day, and quotas were reset for all users.

This is a confirmed operational incident, not an evaluation dispute. It also shows that model capability releases and the shakedown of infrastructure such as quotas and billing often do not complete on the same timeline.

On launch day, Meta's stock rose 3%. The stock reaction is market-level confirmation, but it does not constitute validation of model capabilities.

Several key readings are still pending. Arena has added the models to Agent Arena and Code Arena, but scores are pending.

There are also reservations on capability boundaries: the model is weaker in scientific discovery intuition, and RSI Index results are pending. In addition, whether Muse Wonder belongs to the Fable/Astra class of models is unclear because of differing degrees of post-training; Wonder Full beats Fable and Astra on some benchmarks, but which specific benchmarks is not detailed.

On safety, Meta says the family models were trained for robustness against adversarial instructions and handling of private information, but no detailed system card was provided. Therefore this can only be recorded as a vendor statement and cannot be used to extrapolate independent conclusions.

Putting these gaps together, the certain parts of this release are: the models are live, scores have been measured by third parties, and pricing has been half published. The uncertain parts are: the cost of the full version, some arena rankings, and the true upper bounds of several capability dimensions.