Meta's Muse Wonder reaches frontier tier with Wonderful Full

Muse Wonderful Full places Meta third overall on the Artificial Analysis benchmark, beating all OpenAI models on ARC. The model is the first from a lab outside the usual duopoly to reach the tier of Grok and Astra.

The release splits into two products: Muse Wonder at 120 tokens per second, available via API with disclosed pricing; and Muse Wonderful Full at 50 tokens per second, locked behind Meta's subscription with no pricing revealed.

After days of using Muse Wonder, the model represents a massive jump from Spark 1.3, Meta's previous standard. Benchmark scores reflect genuine capability improvements across the board.

The tester found the model's default behavior pleasant without modification. Removing ADHD-friendly formatting instructions (short responses, varied sentence length, simplified English) made the model more verbose, but it remained effective at coding and reasoning tasks.

As a reviewer in real code work, Muse Wonderful was thorough and correct—it wrote 209 probes against Opus fixes and provided clear grouping of issues, while also serving as a fair judge that ranked its own entry second.

Muse Wonder runs at 120 tokens per second; Wonderful Full operates at 50 tokens per second, indicating a larger underlying model. Only Muse Wonder carries public API pricing; Wonderful Full is restricted to subscription.

The split makes price-to-performance comparison impossible for the capable version. Without published pricing or subscription limits for Wonderful Full, buyers cannot assess whether the model justifies switching from competitors.

On terminal benchmark scores, Wonderful Full destroys Grok 5.1's numbers and approaches Grok 4 and Claude Opus 3.5. It achieves state-of-the-art on Deep Suite and SWE Marathon benchmarks.

API pricing for Wonderful Full is set at $4 per million input tokens, but the subscription model's economics remain opaque.

A third-party company appears to have developed and sold a 3D training system to multiple labs, explaining the synchronized improvements.

Claude 3 Haiku first demonstrated this pattern, then the trend replicated across frontier labs. Opus 3.5 showed massive 3D capability gains despite scoring only 5% higher overall than Fable 5.1, showing that model smartness and 3D ability do not correlate directly.

The same pattern applies to computer use and other specialized capabilities. Frontier labs license post-training environments for particular skills rather than discovering them through general model intelligence. This distinction matters for understanding where lab differentiation actually occurs.

For users who need 3D tasks, Anthropic's $200 monthly cloud code plan delivers exceptional value—superior to maintaining local GPUs, using budget Chinese labs, or buying tokens through open router services.

No comparable pricing structure from Meta exists yet, since Wonderful Full's subscription costs and limits remain unpublished.

The model generates slop comparable to Sonnet 3.7, whereas Grok fails at UI generation in visibly different ways.

This weakness directly impacts developer experience. Anyone pulling the API and integrating it into existing tools will encounter bad performance unless they use Wonderful Full instead.

Using a standard agent markdown prompt with no ADHD-friendly modifications, Muse Wonder became the tester's favorite model to work with immediately. The model's default verbosity and interaction style outperformed other frontier models without sacrificing reasoning ability on most tasks.

The difference traced to removed accessibility instructions in the prompt—the model no longer constrained itself to short responses, varied sentence length, or simplified English, allowing its natural communication style to emerge.

Muse Wonderful fixed test failures caused by incorrect test logic and corrected driver layer issues and variable naming bugs. The model's fixes addressed real problems that required understanding both the immediate issue and the surrounding system.

As a code reviewer, Muse Wonderful proved thorough and accurate. It wrote 209 probes against Opus's fixes and provided clear grouping of issues, catching problems other models missed while ranking its own entry fairly.

Muse Wonderful is the first model from a lab other than Anthropic or OpenAI that the tester would trust to orchestrate sub-agents and workflows for long-running tasks. It represents a third flavor of frontier AI, distinct from both Anthropic and OpenAI approaches.

The model has quirks and performs worse in certain areas—notably UI generation and some benchmarks—but has produced fewer egregious errors than competitors. While the tester acknowledges bias toward novelty (analogous to audio reviewers preferring new speakers simply for being different), Muse Wonderful's actual performance supports treating it as a genuine third option.

Frontier labs now differ not just in scale but in approach. Muse Wonderful's strengths and weaknesses follow a different pattern than Opus or Grok, opening possibility space for specialized teams to prefer it despite lower performance in some benchmarks.

T3 Code's one-click remote machine update feature was failing to update runtime on all machines. The root cause lay in a niche HTTP trace-header bug triggered during npm version checks. Opus correctly identified and fixed the issue but produced verbose, hard-to-parse output.

Muse Wonder delivered highly readable explanation of the trace-header problem and took the analysis further—it discovered an existing pull request from another user that already fixed the issue. None of the other models found the pre-existing fix; Muse alone suggested merging the solution rather than creating a duplicate fix.

Soul missed that the feature was newly added and suggested removing the npm check entirely, the wrong fix. Muse's output combined accuracy with clarity and practical awareness of the existing solution ecosystem.

This bug demonstrates Muse Wonderful's strength at translating complex debugging problems into actionable findings for engineers who need to understand and act on them quickly.

The benchmark tasked models with auditing T3 Code's architecture to identify good, bad, and ugly aspects for prioritization. Opus scored highest with 55 points; Sonnet 3.5 performed surprisingly well; Muse Wonderful Full and Astra landed relatively close.

Standard Muse Wonder dropped to almost half of Opus's score—a more than 30% quality decrease compared to Wonderful Full. The gap between the two Muse versions on this task far exceeds the gap between any single-model options.

Muse Wonder does not appear compelling at its price point. Most online demos showing bad performance are using the API version, integrating it into existing tools without optimization. The standard model's weakness becomes immediately obvious in production use.

Anyone testing the model this way encounters the weaker version's limitations—poor UI, weak architecture analysis, and general underperformance relative to existing frontier alternatives. The two-tier split means many early adopters will experience Muse Wonder and conclude the entire release underperforms.

Muse Wonder averaged 54.2K tokens per task; Muse Wonderful Full averaged 53.8K tokens per task. The near-identical token usage masks a massive capability difference between versions.

The similarity suggests the models may share architecture or training approaches that make scaling through token usage inefficient for Wonderful Full. If true, the models differ not in raw size but in training or deployment specifics—a departure from how other labs scale frontier models.

In code review benchmarks, Opus found the most problems no one else discovered, including four medium-severity issues, but also wrote roughly one nitpick per PR and mistakenly flagged intended behavior as a bug once. Sonnet never wrote a nitpick, meaning every finding mattered, but it found nothing on three of four small PRs—missing real issues even where they existed. Sonnet's unique findings came from deep analysis, like reading upstream source code.

Muse Wonderful had the most overlap with bugs other models found, focusing on issues with clear evidence rather than novel discoveries. It did catch a token revoke issue Opus missed, but its unique findings skewed low severity. Muse Wonderful positioned itself as the practical middle choice—reliable findings with lower false-positive rate than Opus, though less novelty than Sonnet's deep dives.

The codebase analysis task encountered technical issues; it was sent at 9:19 and froze. After resending at 10:24, Muse Wonderful completed at 10:31—a 7-minute response window.

Muse Wonderful's output was more concise and readable than Opus's verbose 30-minute response, delivering actionable architecture analysis in a format easier to parse and act on. Speed and readability combined to create practical advantages for teams reviewing large codebases.

The task highlighted Muse Wonderful's ability to handle extended reasoning on complex systems without producing the explanation bloat that other frontier models introduce. For codebase work, the lean output may justify the model's subscription requirement.

Muse Wonderful identified data-loss issues: destroyed environment files, settings changes that revert, and subscriptions continuing to bill after cancellation. The severity is high, but counts are low and evidence is thin. Each issue requires reproduction on the current nightly build before accepting as confirmed bugs.

Users cannot assess whether Wonderful Full justifies switching from Claude's $200 monthly subscription without knowing Meta's pricing and subscription quality. Meta has not released the price for the model worth using.

Until subscription economics are public, no one can responsibly recommend Muse Wonderful over proven alternatives. The information gap transforms what could be a competitive option into a mystery box.