Five weeks ago, Meta called Muse Spark 1.3 its biggest jump so far on coding and agentic work - a fast, cheap, $1.25-per-million-token model the company said could match Claude Fable 5.1 and beat GPT-5.6 Sol at coding. Today, October 12, 2026, Meta released something priced in an entirely different league: Muse Wonder and Muse Wonder Full, which Meta's Superintelligence Lab is calling its first genuinely frontier-class models. The API rate alone is more than 3x Spark's, and the top Muse Code subscription tier that unlocks Wonder Full runs $200 a month - roughly 160 times what a comparable chunk of Spark usage costs through the API.
That's a big jump in both price and ambition in five weeks, and it raises an obvious question: what does Wonder actually do that Spark 1.3 doesn't, and is the extra cost justified by anything more than a marketing label? Here's what's actually confirmed, what's still Meta's own claim, and how the math on pricing really works out.
Quick summary: Meta released Muse Wonder and Muse Wonder Full on October 12, 2026, describing them as its most capable models yet, built around two new features - Active Memory (retains and synthesizes what you've told it across sessions) and Active Focus (keeps working toward a goal, retries failed approaches, and checks in before consequential actions). Via the Meta Model API, Wonder costs $4 per million input tokens and $20 per million output tokens. Wonder Full isn't sold via API at all - it's exclusive to Muse Code subscribers, on a Power tier at $50/month or a Maximum tier at $200/month. Meta says Wonder scores higher than Fable, Opus, or Astra on its own internal MemoryBench test, and that both Wonder models post benchmark scores comparable to Claude Fable 5.1 and GPT-6 Astra - but Meta hasn't published the actual numbers behind the memory claim, and no independent evaluator has verified it yet. The models were trained at Meta's gigawatt-scale Prometheus data center in central Ohio, which came online earlier this year.
What Active Memory and Active Focus actually mean
Meta is positioning Wonder less as "a smarter chatbot" and more as a model built to handle real multi-step tasks that span conversations - the examples Meta itself gives are things like navigating an insurance claim or understanding a lease, not a single Q&A exchange. Active Memory is the half built for continuity: instead of starting cold every session, Wonder is designed to carry details about you - your situation, your preferences, your ongoing task - forward, and surface the relevant pieces when they matter again.
Active Focus is the half built for persistence inside a single hard task. Meta describes it as the model trying alternative approaches when the first one fails, recovering from errors instead of giving up, asking you for clarification when it's genuinely stuck, and - notably - checking in with you before following through on anything consequential. That last piece matters: it's Meta building in a pause-and-confirm step for exactly the kind of agentic action (booking something, submitting a form, spending money) that's caused real problems for other labs' agents this year.
Where Wonder is meant to be used: expert domains, not just chat
Meta is pitching Wonder and Wonder Full at what it calls expert domains - law, finance, and health - plus what it describes as frontier-level coding performance, aimed at people building anything from a small game to a working app for a new business. That's a deliberate shift upmarket from how Meta has talked about its AI products before: this isn't framed as a free assistant bundled into Instagram or WhatsApp, it's framed as a tool for people doing professional-grade work, priced accordingly.
Pricing: Spark 1.3, Wonder, and Wonder Full side by side
Spark 1.3 and Wonder are sold through two channels that don't map cleanly onto each other - Spark and Wonder (but not Wonder Full) are available through the pay-as-you-go Meta Model API, while Wonder Full only comes bundled into a flat-rate Muse Code subscription. Here's how the confirmed numbers compare:
| Model | Access | Price | Positioning |
|---|---|---|---|
| Muse Spark 1.3 | Meta Model API / Muse Code | $1.25 in / $4.25 out per 1M tokens | Lightweight, low-latency flagship (released Sept 2, 2026) |
| Muse Wonder | Meta Model API | $4 in / $20 out per 1M tokens | New frontier tier - larger, more capable, pricier |
| Muse Wonder Full | Muse Code only (not via API) | $50/mo (Power) or $200/mo (Maximum) | Full agentic capability, flat-rate subscription only |
The API comparison is the cleanest one: Wonder costs 3.2x Spark 1.3's input rate and 4.7x its output rate. The subscription comparison is where the "160x" framing comes from, and it needs a caveat - Muse Code's $200/month Maximum tier isn't metered per token, it's a usage allowance (Meta describes it as roughly 80 times the usage of its base "Everyday" tier), so comparing it directly to Spark's per-token API price only works as a rough, worst-case illustration of how much further upmarket Meta has priced Wonder Full, not an apples-to-apples rate.
If you just want to try Wonder without committing to a subscription, the Meta Model API is the cheaper entry point - you pay per token and can test it against a specific task before deciding whether Wonder Full's subscription-only features (and its $50-$200/month price) are worth it for your use case.
The MemoryBench claim deserves a hedge
The single most attention-grabbing number in Meta's launch materials is that Muse Wonder "scores higher on MemoryBench than Fable, Opus, or Astra" - a direct shot at Anthropic's and OpenAI's current flagships. It's worth being precise about what that claim is and isn't. MemoryBench appears to be Meta's own benchmark, evaluated on Meta's own infrastructure; neither the exact scores nor the methodology have been published in a way that lets Artificial Analysis, llm-stats, or any other independent evaluator reproduce it. Meta's broader claim - that Wonder and Wonder Full post benchmark scores "comparable to" Claude Fable 5.1 and GPT-6 Astra on more general tests - is more modest and more plausible, but it's still Meta grading its own launch-week homework. Until a neutral third party runs its own memory-retention test across all four models, treat the MemoryBench lead as Meta's claim, not a confirmed fact.
Spark 1.3 isn't being replaced - it's being joined
It's worth being clear about what changed and what didn't: Meta isn't retiring Muse Spark 1.3. Spark was built and marketed specifically as a lightweight, low-latency model - fast responses, low cost, strong enough to beat GPT-5.6 Sol on coding benchmarks by Meta's own account. Wonder is a different, larger tier sitting above it: bigger, more expensive, and built for the kind of long-horizon, multi-domain, memory-dependent agentic work Spark was never positioned to handle. Meta now has three distinctly different Muse products aimed at three different jobs - including the smaller, free, open-weight Muse Glimmer - which is confusing enough that it deserves its own breakdown.
- Developers building professional agentic tools in regulated domains (legal, financial, health-adjacent) who need a model Meta is specifically marketing for that use case.
- Teams already paying for Muse Code who want to test whether Wonder Full's persistence and memory features meaningfully reduce the babysitting a long multi-step agent task usually needs.
- Anyone comparing frontier options across labs before committing a monthly budget - Wonder's API pricing ($4/$20 per million tokens) sits close to several of Anthropic's and OpenAI's mid-to-high tiers, so it's a fair one to benchmark against your actual workload rather than against Meta's own claims alone.
It's a weaker pick for casual chat or simple day-to-day questions - Spark 1.3 is cheaper, faster, and was built for exactly that, and Muse Code's free "Everyday" tier exists for lighter use before any subscription decision is needed.
01What is Muse Wonder, exactly?
It's Meta's newest and, per Meta, first truly frontier-class AI model, released October 12, 2026 alongside a larger sibling, Muse Wonder Full. Both are built around Active Memory (retaining context across sessions) and Active Focus (persisting through multi-step tasks and checking in before consequential actions), and are marketed for expert-domain work in law, finance, and health, plus frontier-level coding.
02Is Muse Wonder the same as Muse Spark 1.3?
No. Spark 1.3 (released September 2, 2026) is Meta's lightweight, low-latency flagship, priced at $1.25/$4.25 per million tokens. Wonder is a separate, larger, pricier tier released five weeks later at $4/$20 per million tokens via API, or $50-$200/month as a Muse Code subscription for Wonder Full. Meta is keeping both available rather than replacing one with the other.
03Does Muse Wonder actually beat Claude and GPT-6 at anything verified?
The one specific claim - that it beats Fable, Opus, and Astra on Meta's own MemoryBench test - comes entirely from Meta's own testing, with no published scores or independent verification yet. Meta's broader claim that Wonder's general benchmark scores are "comparable to" Fable 5.1 and GPT-6 Astra is more modest, but still self-reported.
04How much does Muse Wonder actually cost?
Via Meta's API, Wonder is $4 per million input tokens and $20 per million output tokens. Wonder Full isn't sold via API - it only comes through a Muse Code subscription, at $50/month (Power tier) or $200/month (Maximum tier), both of which include access to both Wonder and Wonder Full.
05Should I switch from Spark 1.3 to Wonder?
Only if your workload actually needs what Wonder adds - persistent multi-session memory, expert-domain reasoning, or long multi-step agent tasks. For everyday chat, coding help, or simple queries, Spark 1.3 remains cheaper and, by Meta's own account, nearly as capable for those jobs.
If you want to see how Muse Wonder actually performs next to Claude Fable 5.1, GPT-6 Astra, or Grok on a task you care about - rather than taking Meta's MemoryBench claim on faith - LumiChats gives you one login to chat with 40+ leading AI models side by side, so you can judge a launch-day claim against your own prompts instead of a vendor's own benchmark slide.
