Meta Muse Wonder Released: 56 AA Score and API Pricing

👁 112 reads

Meta has unveiled Muse Wonder, its first frontier model family competing at the very top. In vendor benchmarks, Muse Wonder Full outperforms Claude Opus 5.5, Claude Fable 5.1, and GPT-6 Astra across 7 out of 14 evaluations, scoring 56 points on the independent Artificial Analysis index, just two points behind Opus 5.5. The smaller Muse Wonder is already available via API at $4 per million input tokens and accessible globally through OpenRouter. I evaluated it immediately using Polish prompts. Detailed breakdown in our Muse Wonder guide.

What Meta Announced

The announcement arrived on Monday, October 12, 2026: a blog post from Meta AI Research, a technical model card, and updates from Mark Zuckerberg on X. The release comes from Meta Superintelligence Labs, an initiative built from scratch just over a year ago. Zuckerberg noted that while Muse Spark served as the team’s initial proof of concept, Muse Wonder represents a “substantially larger and more advanced frontier class” that places Meta alongside top-tier AI labs. Key technical highlights:

• Two models: Muse Wonder and the flagship Muse Wonder Full;
• Training: completed at the gigawatt-scale Prometheus data center with pre-training and reinforcement learning;
• 1M token context window: multimodal inputs (text, image, video, audio) with text output (per the model card);
• Two major agentic features: Active Memory (cross-session persistent context) and Active Focus (autonomous task exploration and retrying alternate execution paths);
• Closed weights: no release plans were indicated in the announcement or model card.

Vendor Benchmark Results

Meta evaluated both models against top tier options from Anthropic and OpenAI, with reasoning set to maximum levels across all models. These figures reflect vendor-provided metrics and internal evaluations:

This subset covers 7 of the 14 reported benchmarks, including areas where Meta fell short. The full breakdown is available in our guide. The models show notable strength in sustained coding and long-term memory retrieval, while lagging on ProgramBench and HealthBench Pro compared to Opus 5.5 and Astra. Meta noted in the model card that the SWE-Marathon score for Full was evaluated at the xhigh reasoning setting, meaning it is not a direct apples-to-apples comparison.

Independent Artificial Analysis Rankings

Artificial Analysis, running third-party tests, awarded Muse Wonder Full 56 points on its Intelligence Index 4.3, while standard Muse Wonder scored 52. For comparison, Claude Opus 5.5 holds 58 points, while Claude Fable 5.1, GPT-6 Astra, and Gemini 4 Argon each sit at 53; Meta’s earlier Muse Spark 1.3 reached 48. In the overall index, only Claude Opus 5.5 (max setting) ranks higher at 58, while Claude Sonnet 5.5 (max) and Opus 5.5 (xhigh) match Wonder Full at 56 (Sonnet 5.5 was omitted from Meta’s comparison charts). The standard Wonder ranks 10th overall, alongside GPT-6.1 Sol. The benchmark also noted verbosity: Wonder generated 130M tokens during scoring, while Wonder Full generated 140M.

Pricing and Availability

• Meta Model API: identifier muse-wonder-1.0, priced at $4.00 per million input tokens, $20.00 per million output tokens, and $0.20 per million cached input tokens (matching Claude Opus 5.5 tier, without the lower Contributor discount seen on Muse Spark).
• OpenRouter: model meta/muse-wonder at matching rates, supporting a 1,048,576 token context window.
• Muse Code: the agent coding tool offers Muse Wonder Full strictly on Power Usage ($50/month) and Maximum Usage ($200/month) plans.
• Muse App: Wonder will roll out to paid subscribers ($20 and $100 monthly tiers), currently focused on US and Canada availability.
• Meta AI Integration: Meta stated that Wonder will expand to WhatsApp, Instagram, and Messenger in the future, without specific regional timelines.

Evaluation in Polish: 87.0 Score

Shortly after launch, I ran Muse Wonder through my standard evaluation suite: 13 complex tasks, where we tested it in Polish alongside over 20 prior models, assessed blind by two external evaluators. The calibrated score reached 87.0 points. This positions it in second place on our leaderboard, behind Claude Opus 5.5 (90.1) and ahead of GPT-6 Sol (86.4), GPT-6 Astra (85.3), and Claude Fable 5.1 (83.9). The predecessor, Muse Spark 1.3, scored 75.3. Running this evaluation cost $0.45 in API credits. On the standardized Polish high-school exam prompts (CKE May 2026), Muse Wonder scored 215 out of 220 points. Specific edge cases and failed prompts are detailed in the full guide.

What This Means for Frontier AI

Over the past year, the leading frontier was primarily contested by Anthropic, OpenAI, and Google. Following the underwhelming reception of Llama 4, Meta restructured its AI teams, delivering a competitive model within 14 months that matches Anthropic near the top of the independent AA leaderboard. Developers worldwide have immediate access via the API and OpenRouter at Opus 5.5 pricing tiers. As wider community benchmarks like LMArena publish results, we will see whether Meta maintains its declared lead in autonomous agent tasks.

// read moreMeta Muse Wonder: Benchmarks, API Pricing, and Full TestBenchmarksHow to Use Muse Wonder: OpenRouter, Meta API, and Muse CodeGuidesMeta Muse Wonder Benchmarks Explained: All 14 Tests DeconstructedBenchmarks

Kinetyra