Meta Releases Muse Wonder: Six Reasoning Levels, Performance on Par with GPT-6.1 Sol

Meta AI has launched Muse Wonder and Muse Wonder Full, which the company calls "the strongest models to date." According to Simon Willison's blog, the new models have been integrated into the LLM toolchain via version 0.2 of his llm-meta-ai plugin, and Meta's research blog published an announcement at the same time.

Six Reasoning Levels: From minimal to max

The most distinctive design feature of Muse Wonder is its built-in six tiers of reasoning intensity: minimal, low, medium, high, xhigh, and max. This tiered approach aligns with the industry trend of "allocating compute according to task difficulty"—simple queries run at lower tiers to save cost, while complex reasoning tasks go all out.

But there are clear gaps in the original information: the official announcement does not explain what exactly changes at each tier, nor does it provide differences in cost or latency across tiers. In other words, the six levels currently look more like a product interface than a verifiable performance promise. For developers preparing to integrate the API, this means tier selection can only be figured out through hands-on testing.

"Reasoning intensity tiers" are a common productization approach for reasoning models in recent years. Their technical basis typically comes from test-time compute scaling: the model generates one or more segments of internal thinking (chain-of-thought) before producing the final answer, and the higher the tier, the larger the allowed thinking token budget, number of samples, or search depth. OpenAI's reasoning effort parameter and Anthropic's extended thinking budget follow the same idea. As a result, different tiers often do not change the model weights themselves but instead change the compute investment during the inference phase, directly affecting latency and per-unit pricing. This also explains why vendors need to explicitly disclose the token budget for each tier—otherwise developers cannot estimate their bills.

Third-Party Evaluation: Max Tier on Par with GPT-6.1 Sol

Artificial Analysis's evaluation shows that Muse Wonder at the Max tier performs comparably to GPT-6.1 Sol. This is a noteworthy signal—if the conclusion holds, Meta has entered the top tier for high-end reasoning models.

Willison further speculates that this may be among Meta's first "Fable class" models. It should be made clear that this classification is his personal interpretation, and Meta did not confirm it in the original announcement. The evaluation data comes from a third-party organization, while the model classification is an observer's judgment—the two do not carry the same weight of credibility.

Artificial Analysis is an independent model evaluation organization. Its approach is to call models uniformly through each vendor's API, run them on standardized question sets, and simultaneously measure output speed (tokens/s), first-token latency, and cost per million tokens, ultimately combining these into a composite intelligence index. Because the evaluation is completed in a third-party environment and does not rely on vendor-reported data, its results are often used as a reference for cross-model comparison. But its limitations should be noted: question set coverage is limited, and results drift as vendors silently update their models, so a single ranking is better treated as a trend signal than a final conclusion.

A Monthly Release Cadence—Is the Progress Real?

According to the timeline outlined by Willison, Muse 1.3 was released on September 2, and Muse 1.2 on August 5—Meta is basically maintaining a pace of one new model per month.

His judgment on progress between versions comes from an informal test: asking each model to draw "a pelican riding a bicycle." Willison believes the improvements between versions are substantive. However, this test is only a quick visual check and does not constitute a formal benchmark, and the original source does not provide other metrics for measuring version gains. For readers following the evolution of model capabilities, this conclusion should be treated as directional reference rather than definitive.

"Drawing a pelican riding a bicycle" is an informal visual test circulating in the AI community, which became popular after people used it to examine image models' spatial relationships and object composition abilities. It is used repeatedly because the pelican's beak, the bicycle frame, and the rider's posture create multiple constraints, easily exposing flaws in part assembly and physical plausibility. But such single-prompt tests have a sample size of one, and results depend heavily on random seeds and prompt wording—they are neither reproducible nor quantifiable, and can only serve as qualitative observations, not as substitutes for standardized benchmarks such as MMLU or GPQA.

Information Gaps: Pricing, Context, and Launch Date All Undisclosed

The original source contains no pricing, rate limits, context window size, or official launch date. These are precisely the key parameters for production environment selection. Teams planning to use Muse Wonder in actual business operations need to consult Meta's official documentation directly to fill in these details.

From a product strategy perspective, Meta first signaling through the plugin ecosystem and third-party evaluations and then gradually disclosing commercial details is a common market warm-up path. But for developers, with key parameters missing, the current stage is better suited to small-scale experiments than large-scale migration.

Three Things Worth Watching

• The actual differences among the six reasoning levels have not been disclosed, and the cost of tier selection must be tested firsthand;
• Artificial Analysis's Max-tier evaluation is on par with GPT-6.1 Sol, but the "Fable class" classification is only a personal interpretation;
• Capability improvements under the monthly release cadence are currently supported only by informal tests, lacking public benchmark support.