Meta Launches Muse Wonder, Codenamed Watermelon — Two New 'Frontier' Models That Close the Gap on OpenAI and Anthropic

Meta shipped Muse Wonder and Wonder Full on Oct 12, internally codenamed Watermelon and trained on roughly an order of magnitude more compute than April's Muse Spark. CEO Zuckerberg called it "frontier"; chief AI officer Alexandr Wang said it "rivals the big (boys)." Artificial Analysis ranks it ahead of prior Meta models plus selected OpenAI, Anthropic, and Google models on its benchmarks — and among the most expensive to run.

Meta launched Muse Wonder and Wonder Full on Monday, October 12, marking the most concrete signal yet that the company’s closed-source pivot from the open-weight Llama era is delivering frontier-class results. CEO Mark Zuckerberg announced the release on X, calling Wonder “a much larger and more advanced class of models that brings us to the frontier.” Chief AI officer Alexandr Wang — installed after Meta’s $14 billion Scale AI deal in June 2025 — followed up: Wonder “rivals the big (boys),” the shorthand on his feed for OpenAI and Anthropic. (France 24 / AFP, Business Insider)

The release lands six months after Meta’s last closed-source push — the Muse Spark family in April — and roughly a year after the company effectively conceded that its open-weight Llama line had fallen behind. The shift from open to closed happened fast: Llama 3-influenced architectures were still being shipped in late 2024, the company restructured around closed Muse releases in early 2025, and now the second-generation closed models are claiming benchmark parity with the labs that spent 2024–2025 building the frontier.

What shipped

Two models, two positioning curves:

  • Muse Wonder — a “smaller, fast” variant Meta expects people to use for daily work. Positioned as the default in the consumer Muse agent app for paid subscribers.
  • Wonder Full — the “larger and more capable” version, built for complex multi-step tasks. Used as the escalation tier when Wonder isn’t enough. (Business Insider)

Both are designed around memory and alternative-approach retry — features Meta highlighted specifically through a “disputing an insurance claim” use case. Multi-step agentic work is the explicit framing; one-shot prompting is the implicit baseline. Memory means state across calls; alternative-approach retry means the model can backtrack and try a different decomposition when one path stalls.

Watermelon, Avocado, and the compute story

The internal codename is Watermelon, per Business Insider’s source. In July, Wang told staff that Watermelon had already caught up to OpenAI’s leading models and was using “an order of magnitude more compute than Avocado” — the codename for Muse Spark. (Business Insider, Business Insider (July))

The 10× compute delta between Avocado (Spark) and Watermelon (Wonder) is the operational story underneath the launch. Meta has guided to roughly $145 billion in AI-related capex for 2026; the company spent more on AI infrastructure in the first three quarters of this year than it spent on the entire Reality Labs metaverse push across any prior year. Wang’s claim of order-of-magnitude compute scaling on a six-month cadence is the consumption pattern that justifies the capex line — and the basis on which investors are being asked to underwrite the next leg. (France 24 / AFP)

The benchmark claim

Meta’s blog post points to third-party benchmarking from Artificial Analysis — an independent firm that publishes model rankings — showing Wonder outperforms:

  • Meta’s own prior Muse Spark family
  • Selected models from OpenAI, Anthropic, and Google (France 24 / AFP)

That last clause is doing a lot of work. “Selected” is the operative word: Artificial Analysis’s published ranking tables break out performance by capability dimension (reasoning, code, long-context, multimodal, agentic), and Meta is highlighting the subset of dimensions where Wonder lands on top. The fairer read is that Wonder is competitive across most capability dimensions rather than uniformly superior.

The same Artificial Analysis data shows Wonder is “among the most expensive models” on the per-token pricing comparison. Frontier-class closed models from OpenAI (GPT-6) and Anthropic (Claude Opus 5) carry similar premium pricing; the competitive question is how much of the price is compute pass-through versus platform-margin capture. Meta’s 3.6 billion daily active users give it a distribution asymmetry no other frontier lab can match — the marginal inference cost of putting Wonder behind every WhatsApp or Instagram interaction is genuinely small, even if the absolute cost is large. (France 24 / AFP)

The full-stack distribution bet

The model launch is the showcase; the distribution stack around it is what’s actually new for Meta’s AI strategy:

  • Muse Spark (April 2026) — first closed-source family, the Avocado generation. Established the pivot from Llama.
  • Muse Wonder + Wonder Full (October 2026) — second closed-source family, the Watermelon generation. Today’s launch.
  • Muse Code (August 2026) — agentic coding tool, direct competitor to Cursor / Claude Code / Copilot. (France 24 / AFP)
  • Muse (September 2026) — agentic consumer app, paid subscription tier. First distribution surface for Wonder.
  • Muse Charm (Connect, September 2026) — handheld AI hardware for interacting with the Muse agent separate from a smartphone. (France 24 / AFP)

Five products in seven months. The model is the headline, but the strategic bet is that Meta can ship a vertically integrated stack — model, agent app, coding tool, hardware — that no other frontier lab currently matches end-to-end. OpenAI has the ChatGPT consumer surface and Codex CLI; Anthropic has Claude Code and the enterprise API; neither has shipped a hardware device. Google’s distribution surface (Search, Android, Workspace) is broader but Gemini’s agent surfaces are less mature. Meta’s bet is that closing the gap on model quality while pulling ahead on distribution surface creates a defensible position without needing a third-party platform.

The safety-shaped delay

Zuckerberg publicly disclosed in September that Meta delayed shipping Muse for several months to focus on safety and security, and called out the industry contrast explicitly: “We didn’t call for everyone else to do this before we would.” (France 24 / AFP)

The framing is a deliberate counterweight to Meta’s “move fast and break things” heritage. Whether the delay was six months or longer isn’t public, but the message is targeted: frontier model deployment at the Wonder scale has catastrophic-failure modes (jailbreaks, agent-loop DoS, exfiltration via multi-step tasks), and Meta is signaling that it took the responsible-disclosure lane rather than the ship-first-and-patch cycle. The signaling targets both regulators (whose posture on agentic AI is hardening) and enterprise buyers (whose procurement gates increasingly require documented safety review).

What this means for the frontier race

Three structural shifts are now legible on the same day:

  1. Meta is no longer behind. The “Meta is years behind on frontier” story — true through Llama 3 — has closed inside 18 months of the closed-source pivot. The next-cycle question is whether Meta can sustain 10×-compute-per-generation cadence or whether Wonder Full is the peak before amortization catches up.
  2. Open-weight has lost a major backer. Meta was the highest-profile advocate for open-weight foundation models through 2024. Closed-source Muse is now the published strategy. The open-weight frontier is Mistral, Alibaba/Qwen, the Chinese open-weight community, and a handful of academic consortium efforts. That’s a meaningfully smaller coalition than 18 months ago.
  3. Compute capex is the binding constraint, not model architecture. Watermelon’s 10× compute scaling over Avocado was the explicit justification for Meta’s $145B 2026 capex line. If the next frontier jump also demands an order-of-magnitude more compute, the gap between labs with hyperscaler-class infrastructure (OpenAI / Microsoft, Anthropic / Amazon, Meta, Google, xAI) and everyone else becomes permanent.

The empirical question for the rest of 2026 is whether Wonder Full holds up under independent third-party evaluation — specifically on agentic long-horizon tasks where the “disputing an insurance claim” framing implies the model is targeted. If it does, Meta joins the four-lab frontier club as a peer rather than a follower.

Sources