Meta finally has an answer for everyone who asked what its $130 billion in AI spending would buy. On Monday, October 12, the company unveiled Muse Wonder and Muse Wonder Full, which it calls its first frontier models, built by Meta Superintelligence Labs. The stock jumped about 3.4 percent on the news.
The launch matters for two reasons. First, independent benchmarks suggest Meta has genuinely joined the frontier group, even if it is not leading it. Second, Meta got there by abandoning the open-model philosophy it spent years championing. The weights are closed, the pricing matches Claude Opus exactly, and the strongest version is locked behind a subscription.
Two models, but only one is for developers
There is a catch buried in the launch announcement, and it is a big one. The standard Muse Wonder is available through the Meta Model API and OpenRouter for developers. The stronger Muse Wonder Full, which Meta says is the more capable of the two, is reserved for paying Muse Code subscribers. It is not served through the API at all.
This is a deliberate product decision. Meta wants developers building on Muse Wonder, but it wants its best model pulling people into its paid coding product and, eventually, its Muse personal agent app. VentureBeat noted the asymmetry bluntly: the weaker model goes to the API, the frontier one stays inside Meta’s subscription walls. Both versions will eventually power the Muse personal agent, which Meta rolled out in September, with the initial rollout going to subscribers first.
The open-source era is officially over
Remember when Meta insisted the future of AI was open weights? Llama defined the company’s AI identity for years. Muse Wonder ships with no downloadable weights, a clean break from that era. This is now the same playbook as OpenAI and Anthropic: API-only access, subscription upsell, closed development.
The pricing tells the same story. Muse Wonder costs $4 per million input tokens and $20 per million output tokens, with cached inputs at $0.20 per million. That is, to the cent, the same list price as Claude Opus 5.5, which we covered when Anthropic launched it in September. For Indian developers, that works out to roughly Rs 360 per million input tokens and Rs 1,800 per million output tokens.
| Model | Input (per M tokens) | Output (per M tokens) | Weights |
|---|---|---|---|
| Muse Wonder | $4 | $20 | Closed |
| Claude Opus 5.5 | $4 | $20 | Closed |
| Muse Wonder Full | Not on API | Not on API | Closed |
When Meta charges exactly what Anthropic charges for its flagship, it is not trying to undercut the market. It is telling the market it belongs in it. Meta’s own benchmarking claims Muse Wonder is comparable to Claude Fable 5.1 and GPT-6 Astra, and Zuckerberg wrote that the release “brings us to the frontier.”
Independent benchmarks: frontier member, not frontier leader
Meta’s own numbers deserve skepticism, but the independent testing looks credible. According to Artificial Analysis, Muse Wonder Full scores 56 on the Intelligence Index, tied with Claude Sonnet 5.5, ahead of GPT-6 Astra, Claude Fable 5.1, and Gemini 4 Argon at 53, but behind Claude Opus 5.5 at 58. The standard Muse Wonder scores 52.
Coding is clearly Meta’s sharpest edge. Wonder Full scores 68 on the Coding Agents Index, tying Claude Code powered by Sonnet 5.5. Meta claims its SWE-Marathon result of 57.5 percent beats Claude Opus 5.5’s 46.3 percent, and the model doubled its Terminal-Bench score while using 30 percent fewer output tokens. The weak spots: factual knowledge is soft (its AA-Omniscience score of 30 trails Opus 5.5’s 46), though its hallucination rate of 23 percent is reportedly the second-best among top models.
That profile, strong on long-running agentic coding work, weaker on trivia, matches what Meta is selling. The headline features are “active memory,” which carries context across conversations for days, weeks, or months, and “active focus,” which Meta says keeps the model working persistently on hard tasks. Meta research scientist Kavya Kopparapu said the team put its effort into memory “over days, weeks, months, and beyond.” Artificial Analysis’ MemoryBench results rank it above Fable, Opus, and Astra on memory tasks, which is at least consistent with Meta’s own story.
What this means for the race
Three things stand out. First, Meta Superintelligence Labs, the division built around Alexandr Wang after Meta’s $14 billion Scale AI deal, has produced its first genuinely competitive model about a year after the division was formed. That is fast by the standards of frontier labs.
Second, Wall Street is buying the story. Meta stock is up roughly 31 percent since the start of September, and the Muse agent app’s quick adoption drove a 26.7 percent gain in September alone. BofA analyst Justin Post wrote that growing evidence of MSL investments translating into model performance could keep supporting the stock. When you are guiding $130 to $145 billion in capital expenditure for 2026, you need the market to believe in the product, not just the servers.
Third, and most interesting: Meta’s research blog reportedly still has no technical writeup for the model while the product surface is already live on OpenRouter. A frontier lab that ships product first and publishes later looks a lot like OpenAI circa 2024. Meta spent years positioning itself as the open, research-first alternative. Muse Wonder is the moment it stopped being that and started being a competitor instead.
Sources: Meta newsroom announcement; Bloomberg; Barron’s; Reuters; TechCrunch; VentureBeat; The Decoder; Artificial Analysis. Also read: Claude Opus 5.5 Arrives: Anthropic’s New Flagship Costs 40% Less and Codes Faster; Claude Filed a False Police Tip and 20 Visa Applications; DeepSeek Is Raising Up to $15 Billion as the AI Funding Frenzy Hits Full Speed.
