Muse Wonder API Guide: Pricing, Access, and What Changed from Spark
How developers get Muse Wonder on the Meta Model API: $4/$20 per million tokens, preview access rules, what's different from Spark 1.3, and when Wonder Full is
By Professor Muse7 min read
Developers can access Muse Wonder through the Meta Model API at $4 per million input tokens and $20 per million output tokens, with a 1M-token context window, multimodal inputs, and built-in tool calling for agentic work. Only the standard Wonder tier is API-served: Wonder Full is gated behind Muse Code subscriptions. Spark 1.3 remains the cheaper option at $1.25/$4.25.
Key takeaways
- Muse Wonder is in preview on the Meta Model API as
muse-wonder-1.0; third-party guides also list it on OpenRouter asmeta/muse-wonder. - Pricing is $4 in / $20 out per million tokens, with reported cached input at $0.20 per million and a web-search add-on of $2.50 per 1,000 queries.
- Muse Wonder Full has no API route: it ships only inside Muse Code’s $50/month Power Usage and $200/month Maximum Usage plans.
- Context is about 1M tokens with text, image, video, and audio inputs and text output; the model supports tool calling, structured output, search grounding, and computer use.
- Reported default rate limits are 3,000 requests and 4M tokens per minute per team, but treat these as community-reported until Meta’s docs confirm your tier.
Getting access
Meta launched Wonder in preview on October 12, 2026, the same day Mark Zuckerberg announced it, and the API path was live immediately. The consumer Muse app rollout comes later, with paying subscribers first, so right now the API and Muse Code are the two ways to use it.
Getting an API key goes through Meta’s developer platform. Meta’s developer documentation labels Wonder as “in preview,” which usually means features and limits can shift between now and general availability. The practical flow, per independent developer guides published on launch day:
- Create or sign into a Meta developer account.
- Accept the model usage terms for the Meta Model API.
- Generate an API key in the dashboard and attach a billing method for pay-as-you-go usage.
- Point your client at the
muse-wonder-1.0model ID.
One note of caution: the finer-grained details below (exact base URL, request formats, reported rate limits) come from third-party developer write-ups published at launch, not from a line-by-line check of Meta’s live docs. They look right and agree with each other, but verify against Meta’s current developer documentation before you ship anything production on them. Meta’s own announcement is the canonical place to start; our pricing breakdown covers what the numbers mean for your bill.
Pricing: API vs. app
| Path | Model | Cost |
|---|---|---|
| Meta Model API (pay-as-you-go) | Muse Wonder | $4 / 1M input tokens, $20 / 1M output tokens |
| Meta Model API (cached input, reported) | Muse Wonder | $0.20 / 1M cached input tokens |
| Meta Model API (add-on, reported) | Muse Wonder | $2.50 / 1,000 web searches |
| Muse Code subscription | Muse Wonder + Wonder Full | $50/mo Power Usage; $200/mo Maximum Usage |
Cached input at a reported $0.20 per million makes repeated agent loops much cheaper: anything your agent re-reads across turns, like a codebase context or a long brief, is priced at a 95% discount relative to fresh input. Web search is billed separately when you use search grounding, so factor that into agent designs that browse often.
The headline comparison: Wonder’s list price matches Claude Opus 5.5 exactly ($4 in, $20 out, and the same reported $0.20 cache rate), per the launch-day comparison from dev.to. GPT-6 Astra lists at $10 in and $50 out, so Wonder undercuts that tier by more than half. Wonder is roughly 3.2x the input price and 4.7x the output price compared with Spark 1.3’s $1.25/$4.25, which means Wonder has to earn its premium in fewer retries and fewer wasted tokens.
What changed from Spark 1.3
Spark was built for speed: a lightweight, low-latency model that made agent loops cheaper per call. Spark 1.3 refined that story with fewer tool calls and fewer tokens per task (about 20% fewer tool calls and 25% fewer tokens, per Meta’s own framing) plus an experimental “Contemplating” mode that runs sub-agents in parallel to cross-check logic.
Wonder moves the other direction. It is a larger frontier-class model, slower and more expensive, aimed at capability rather than latency:
- On the Artificial Analysis Intelligence Index, Muse Wonder scores 52 versus Spark 1.3’s 48, and Wonder Full scores 56, in the top three on six of the index’s ten tests. Wonder Full ranks #1 on physics reasoning and professional document work.
- Meta’s big differentiator is Active Memory: the model’s ability to retain what matters across long conversations and, where the host app supports it, across sessions. Meta’s MemoryBench hides facts in 64K-100K token conversations, changes some mid-stream, and both Wonder versions score about 85.5, above everything else Meta tested. Active Memory is baked into training; there is no switch to toggle.
- Active Focus is the other headline concept: the model is trained to keep trying when things fail, replan, work around obstacles, and check in before doing anything important. Meta’s own model card recommends requiring human confirmation before irreversible actions, which is sound advice for any agent you put on Wonder.
- Wonder exposes reasoning levels (reported as minimal through max), and Wonder Full is distinguished in Meta’s docs from even the standard Wonder model’s highest reasoning setting, meaning the top-end capability genuinely requires the Muse Code subscription.
The honest tradeoff: Spark 1.3 is still the economical choice for fast, high-volume agent work where latency matters. Wonder is for the tasks Spark couldn’t close: complex coding, multi-app workflows, and long-horizon professional work.
Practical notes for developers
Match the model to the workload. Use Wonder where agent success rate matters more than per-token cost: complex coding tasks, multi-step research, document-heavy professional work. Keep Spark 1.3 (or cached-input patterns) for high-frequency, low-complexity agent loops. Independent evaluators recommend running 30-50 of your own real tasks side by side before switching anything.
Design around output-token economics. Wonder’s 5:1 output-to-input price ratio is the steepest part of the bill. Artificial Analysis found Wonder Full used roughly 140M output tokens to run its index against a median of 88M across models, so reasoning effort settings directly translate to spend. Start at a lower reasoning level and escalate only for tasks that fail.
Mind the privacy surface. Memory is Wonder’s selling point and its risk. An assistant that remembers everything retains a lot about a user, so check Meta’s data terms for how prompts are used, and be deliberate about what your app sends. The model card reports strong alignment results on agentic-misalignment testing, but the persistent-memory properties deserve their own review in your threat model.
Wonder Full’s gating is a real limitation. If you want Meta’s best model inside your own product, you currently cannot have it. Wonder Full exists only inside Muse Code, and Meta has said nothing about an API route or open weights. Plan your roadmap around muse-wonder-1.0, whose capabilities are available to build on.
API formats are familiar. Independent guides report Wonder works through OpenAI-compatible Responses and Chat Completions formats as well as Anthropic Messages, at https://api.meta.ai/v1. That means most teams can trial it by swapping a model ID and base URL rather than rewriting integration code, though you should confirm the current docs before relying on this.
FAQ
How do I get API access to Muse Wonder? Sign up for a Meta developer account, accept the model usage terms for the Meta Model API, generate an API key, and enable pay-as-you-go billing. Wonder is in preview, so expect terms and limits to evolve.
What is the model ID for Muse Wonder?
muse-wonder-1.0 on the Meta Model API, and meta/muse-wonder on OpenRouter, per launch-day developer guides.
How much does the Muse Wonder API cost? $4 per million input tokens and $20 per million output tokens. Cached input is reported at $0.20 per million, and web search grounding at $2.50 per 1,000 queries. Meta charges no long-context premium.
Can I use Muse Wonder Full through the API? No. Wonder Full is available only through Muse Code subscriptions: $50/month Power Usage or $200/month Maximum Usage. There is no public API price or route for it.
What are the rate limits? Community guides report 3,000 requests and 4M tokens per minute per team. These are not yet confirmed in Meta’s official docs, so check your dashboard for your actual tier.
What changed from Muse Spark 1.3? Wonder is a larger, slower, more capable frontier model: about 4 points higher on the Artificial Analysis index (52 vs 48), with trained-in Active Memory and Active Focus for long agentic work. Spark remains cheaper ($1.25/$4.25) and better suited to fast, high-volume tasks.
Is Muse Wonder open source? No. Like Spark, it is closed: no published weights, available only through Meta’s API, Muse Code, and Meta’s apps.
When does Wonder reach the Muse consumer app? Meta has said paying subscribers get it first, with no date announced for free users.
