Meta's Muse Wonder and Wonderful: What They Cost and Who Can Use Them
Meta's closed Muse Wonder and Wonderful models rival Claude Opus 5.5. Here's the pricing, access restrictions, and why Wonderful has no API.
What are Muse Wonder and Wonderful?
Muse Wonder and Muse Wonderful are two closed frontier models from Meta’s superintelligence lab, released within days of the open-weights Muse Spark Open model. Unlike Spark Open, neither Wonder nor Wonderful ships with downloadable weights. Meta has benchmarked them directly against GPT-6 Astra and Claude Opus 5.5, and on the Artificial Analysis Intelligence Index, Wonder scores 52 and Wonderful scores 56, putting Wonderful just two points behind Opus 5.5’s 58. That’s a serious showing from a lab that was only created a little over a year ago.
TL;DR
- Muse Wonder is available through Meta’s own API and through OpenRouter at $4 per million input tokens and $20 per million output tokens, the same price as Claude Opus 5.5.
- Muse Wonderful has no public API at all. It’s locked inside Meta’s Muse Code product, accessible only on the $50 or $200 per month subscription plans.
- Meta’s open model, Muse Spark Open, is a separate, smaller-scoring model (39 on the Artificial Analysis Index) and is not the same model powering Meta’s consumer Muse personal-agent app.
- On OpenRouter, only Meta itself hosts Muse Wonder, while Spark Open already has multiple providers (Modal, Together AI, Nebius, Fireworks, Base Ten) competing on identical pricing.
- A cheaper, data-training-discounted version of Spark 1.3 exists at roughly 10 cents in and 20 cents out, but it requires letting Meta train on your prompts and outputs.
- Despite strong benchmark scores, Wonder ranked 12th on LM Arena’s agent leaderboard after more than 5,000 real agent sessions, even though it placed second on the text leaderboard.
- Meta’s actual breakout product, the Muse personal-agent app, runs on a custom fine-tune of Spark, not on Wonder or Wonderful, and has separate consumer pricing entirely.
How much does Muse Wonder cost through the API?
Muse Wonder is priced at $4 per million input tokens and $20 per million output tokens, whether you access it through Meta’s own API or through OpenRouter. That pricing lines up exactly with Claude Opus 5.5, which signals Meta is positioning Wonder as a direct competitor at the high end rather than undercutting rivals on cost. On OpenRouter, Meta is currently the only listed provider for Wonder, a sharp contrast to Spark Open, which launched day one with half a dozen inference providers all charging the same $1.25 in / $4.25 out rate. That lack of competing hosts for Wonder means there’s no price pressure yet, and anyone testing it should also note that Meta retains prompts sent to the model even while stating it doesn’t use them for training.
Why isn’t Muse Wonderful available through any API?
Wonderful, despite scoring higher than Wonder on the Artificial Analysis Index (56 versus 52), isn’t available through Meta’s API, OpenRouter, or any third-party host. The only way to use it is inside Muse Code, Meta’s coding product, and only on the $50 or $200 per month subscription tiers. There’s no pay-as-you-go access and no way to benchmark it independently through standard tooling. That’s a notable gap for a model Meta is actively comparing to Opus 5.5 and GPT-6 Astra in its own marketing. Until Meta opens an API for it, or a third party gets licensed access, outside verification of Wonderful’s claimed capabilities will stay limited to whatever Meta chooses to publish and whatever shows up on arena-style leaderboards that include it.
How does Muse Spark Open fit into the lineup?
Spark Open is a mixture-of-experts model with 816 billion total parameters and 42 billion active per token, running a 1 million token context window and natively handling text, images, video, and audio as input (output is text only). It scores 39 on the Artificial Analysis Intelligence Index, well behind Wonder (52), Wonderful (56), and even Meta’s own mid-tier closed model, Muse Spark 1.3, which scores 48. That gap matters because many people expected Spark 1.3 to be the version Meta open-sourced. Instead, the open release sits nine points behind that mid-tier closed model and seventeen points behind Wonderful.
Among open models generally, Spark Open isn’t the leader either. Several Chinese open models score higher on the same index, including Mimo 2.6 Pro at 46, GLM 5.3 at 45, Kimi K3 at 44, and Qwen 3.8 at 40, just above Spark Open. Meta’s own framing calls Spark Open “the most capable US open weights model,” a carefully worded claim that concedes the top spot globally while still claiming the US lane.
Licensing is another distinction worth knowing before building on it. Spark Open ships under Meta’s own Muse license, which requires displaying a “built with Muse Spark Open” credit, and anyone running it as a model-as-a-service business with more than $20 million in annual revenue needs a separate commercial agreement with Meta.
Is Muse Wonder worth using over Opus 5.5?
Built like a system. Not vibe-coded.
Remy manages the project — every layer architected, not stitched together at the last second.
On raw benchmark score, Wonder is close to Opus 5.5 but not quite there, and it costs the same per token. That alone makes it a reasonable model to test for teams already comparing frontier providers, especially since it’s accessible through standard API channels without a subscription. But the benchmark picture gets murkier for agentic work specifically. On LM Arena’s agent leaderboard, built from more than 5,000 real agent sessions, Wonder landed at number 12, despite ranking second on the general text leaderboard. That’s a meaningful signal for anyone planning to use Wonder for tool use, multi-step tasks, or autonomous workflows rather than straightforward chat or reasoning tasks. Strong general intelligence scores don’t automatically translate into strong agent performance, and Wonder’s arena ranking suggests a real gap between the two.
Does Meta’s personal agent app run on Wonder or Wonderful?
No. Meta’s Muse personal-agent app, which launched in the US and quickly became the top app on the Apple App Store with more than 2.5 million downloads in its first couple of weeks, runs on a customized version of Spark, not on Wonder or Wonderful. Meta hasn’t disclosed exactly which fine-tune or checkpoint powers the app. According to comments from Meta’s Alexander Wang, the first version of that agent to pass every row in an internal evaluation spreadsheet (covering more than 100 required behaviors) was a specifically trained version of Spark 1.3. The app itself has its own pricing structure, separate from API access: a free tier, a $20 per month tier, and a $100 per month tier, each giving users an isolated virtual machine running their agent and data.
Frequently Asked Questions
What’s the difference between Muse Wonder and Muse Wonderful?
Wonder is the lower-scoring of the two closed models (52 on the Artificial Analysis Index) and is available via Meta’s API and OpenRouter at $4 in / $20 out per million tokens. Wonderful scores higher (56) but has no API access at all, available only through Muse Code’s $50 or $200 monthly plans.
Can I self-host any of Meta’s new models?
Only Spark Open, the open-weights release. It requires substantial hardware (Meta and Apple demonstrated it on two M5 Ultra Mac Studios with 512GB each, running around 25 tokens per second) or quantized checkpoints, including an NVFP4 version around 530GB and a community 2-bit checkpoint around 300GB from Unsloth.
Is there a cheaper way to use Meta’s models via API?
Yes. A discounted version of Spark 1.3 is priced around 10 cents per million input tokens and 20 cents per million output tokens, but using it means allowing Meta to train on your prompts and outputs.
Why does Wonder rank lower on agent benchmarks than text benchmarks?
LM Arena ran Wonder through more than 5,000 real agent sessions and it placed 12th, despite ranking second on the text-only leaderboard. This suggests the model’s general reasoning strength doesn’t fully carry over to multi-step agentic tasks, where tool use and reliability matter more than raw output quality.
Does Meta’s consumer Muse app use any of these new models?
No. The Muse personal-agent app runs on a custom fine-tuned version of Spark built specifically for that product, not on Wonder, Wonderful, or the publicly released Spark Open model.
