Meta releases Muse Wonder and Muse Wonder Full, surpassing Claude on long-horizon coding tests…

[[**Meta's new flagship model Muse Wonder surpasses Claude Opus 5.5 in Agent tests such as long-horizon coding, with its overall intelligence index approaching the top tier, but inference costs surge 183%, and its real-world 3D generation performance is poor, exposing shortcomings in visual-spatial reasoning.**

**Key Points**

**1. Flagship benchmark scores break into the top tier, leading in long-horizon coding**

Wonder Full scores 56 on the overall intelligence index, just 2 points behind Claude Opus 5.5; on the DeepSWE 1.1 long-horizon software engineering test it scores 79.0, surpassing Opus 5.5's 74.2 and GPT-6 Astra's 74.1, placing in the top three on 6 of 10 tests.

**2. Switches to a new pretraining base, optimized for Agent capabilities**

Wonder's internal codename is Watermelon; it uses an independent, larger-scale pretraining run and introduces Active Memory and Active Focus mechanisms to enhance the model's long-term memory, sustained execution, and task generalization from the ground up, rather than relying solely on subsequent reinforcement learning iterations like the Spark series.

**3. Inference costs rise sharply, disruptive pricing strategy draws scrutiny**

Wonder's average task cost reaches $4.52, up 183% from the previous generation Spark 1.3 and 39% higher than GPT-6 Astra; although the official per-token billing price was slightly lowered, the abnormal consumption of subscription quotas for the Full version has been confirmed after a fix.

**4. 3D generation fails in real-world testing, officials acknowledge the shortcoming**

Comparison videos show that the 3D lawn-mower game generated by Wonder Full is far inferior to Opus 5.5 in visual detail, jokingly described as the gap between AAA and Roblox; Meta's model card acknowledges that 3D tasks are a known limitation, and Alexandr Wang responded that the team did not specifically train for such tasks. 2026-10-13 10:10 Has Zuckerberg won AI's comeback battle? A surprise new model takes first place in long-horizon coding, but its 3D real-world test suffers a embarrassing defeat | AI Frontline Quick Look

After being repeatedly left behind by OpenAI and Anthropic, Zuckerberg finally has the confidence to declare: Meta has returned to the frontier of AI technology.

At dawn Beijing time on October 13, Meta Superintelligence Labs (MSL) officially launched its new-generation flagship model series Muse Wonder and Muse Wonder Full, directly challenging top models such as GPT-6 Astra and Claude Opus 5.5.

Data released simultaneously by third-party evaluation agency Artificial Analysis shows that Muse Wonder Full scores 56 on the overall intelligence index, just 2 points behind the current leader Claude Opus 5.5 at 58, and is in the same score bracket as Claude Sonnet 5.5. In multiple long-horizon coding and Agent task tests, it even surpasses existing top models.

On X, Zuckerberg said this is Meta's larger and more advanced generation of models, meaning the company has finally reached the AI technology frontier.

In an episode of the Cheeky Pint podcast released the same day, Zuckerberg and Meta Chief AI Officer Alexandr Wang also rarely discussed the investment behind this comeback.

Meta's core team developing frontier models is only about 100 people, but the computing infrastructure supporting their training is extremely expensive. Zuckerberg revealed that building a 1GW computing cluster costs as much as $50 billion to $60 billion, with annual operating costs of about $15 billion. The scale of capital expenditure Meta discussed at the end of last year even reached the order of $100 billion to $200 billion.

At the same time, Wang also discussed in detail for the first time that Wonder's underlying model has the internal codename Watermelon, used an independently completed, larger-scale pretraining run, and how the team introduced memory and sustained execution capabilities into model training.

But as benchmark scores catch up with the top tier, inference costs have also risen. Third-party estimates show that the average evaluation task cost of the standard Muse Wonder has risen from $1.60 in the previous generation to $4.52, an increase of 183%, and about 39% higher than GPT-6 Astra.

The first developers' real-world experiences are also divided. Some believe Meta already has coding capabilities that can challenge Claude, while others present same-prompt comparison videos questioning that Wonder's performance in 3D program generation is far inferior to Opus 5.5.

Meta finally breaks into the top tier

After the new models were released, Artificial Analysis published the latest tests: Muse Wonder Full scores 56 on Intelligence Index v4.3, and the standard Muse Wonder scores 52.

Compared with Muse Spark 1.3 released six weeks ago, which scored 48, this means the new flagship Full version achieved an 8-point jump, and the standard version also improved by 4 points.

By comparison, Claude Opus 5.5 scores 58, Claude Sonnet 5.5 scores 56, and GPT-6 Astra and Claude Fable 5.1 both score 53.

Looking at the overall results, Meta has indeed returned to the leading camp of frontier models.

In terms of the models' capability distribution, Meta published 14 major evaluations covering personal agents, professional knowledge work, coding, and general intelligence, with competitors including GPT-6 Astra, Claude Fable 5.1, and Claude Opus 5.5.

Among them, Muse Wonder Full scored 79.0 on the DeepSWE 1.1 long-horizon software engineering test, surpassing Claude Opus 5.5's 74.2 and GPT-6 Astra's 74.1.

The DeepSWE test includes 113 tasks from 91 code repositories, covering 5 programming languages. What it tests is not simply writing a function, but whether the model can understand a complex codebase, modify an existing project, and ultimately pass functional and regression tests.

In SWE-Marathon, which places greater emphasis on ultra-long task execution, Muse Wonder Full achieved 57.5%, surpassing Opus 5.5's 46.3% and GPT-6 Astra's 42.5%.

In professional work, Muse Wonder Full also demonstrated competitiveness.

On the legal agent test Harvey LAB-AA, it scored 13.1, surpassing Opus 5.5's 4.2; on the financial research test Vals Finance Agent V2, it scored 59.1, better than Opus 5.5's 58.6.

On its internal MemoryBench test, it scored 85.5, while the standard Muse Wonder was even slightly higher at 85.6, both exceeding Opus 5.5's 81.5.

However, Meta also failed to achieve across-the-board leadership.

On Terminal-Bench 4.0, Wonder Full's 57.1 still lags behind Opus 5.5's 59.6; on ProgramBench, which involves building programs from scratch, Meta scored 59.0, clearly behind Opus's 70.5.

On the computer operation test OSWorld V2.1, Wonder Full's partial-reward score was 78.4, also lower than Opus's 81.8.

In other words, Meta's most prominent advantages are long-horizon software engineering, memory, and some professional agent work, rather than all coding, reasoning, and computer operation tasks.

Third-party Artificial Analysis's Coding Agents Index also provides another piece of supporting evidence: Wonder Full paired with Muse Code scored 68, on par with Claude Code using Sonnet 5.5 and higher than the 66 achieved with Opus 5.5.

Particularly noteworthy is that among the 10 tests included in Artificial Analysis's composite index, Wonder Full ranked in the top three on 6. Among them, it ranked first on the scientific reasoning test CritPt and the PDF-material-based professional work test GDP.pdf.

Compared with the performance leaderboards, the R&D details revealed in the Cheeky Pint interview may be more worth attention. Zuckerberg explained that the Muse Spark series has the internal codename Avocado, and Spark 1.0 through 1.3 use the same pretraining base, with subsequent iterations relying mainly on continual reinforcement learning.

Muse Wonder is different. Its internal codename is Watermelon, and it used a separate, larger-scale pretraining run completed independently.

Wang revealed that Wonder began planning as early as August 2025, rather than being a temporary decision to expand after this year's Spark release. Meta hoped to simultaneously scale up pretraining and reinforcement learning investment, rather than merely continuously stacking post-training on top of an existing model.

Specifically, this technical upgrade is noteworthy in three aspects.

First, a stronger base model and task generalization capability.

In the pretraining stage, in addition to expanding data and compute scale, Meta also placed special emphasis on data from images, video, audio, and professional domains such as law, finance, and health.

Wang pointed out in the interview that to give a model sufficient knowledge in professional scenarios, relying solely on subsequent reinforcement learning is not ideal; the relevant capabilities should begin accumulating from the pretraining stage.

Meta's research blog also provided a set of generalization experiments: comparing Spark 1.3 and Wonder Full under different inference compute budgets, the latter not only improved on tasks targeted by training, but also showed more pronounced capability growth on unfamiliar tasks not directly covered by post-training.

In other words, a stronger pretraining base makes capabilities learned in post-training easier to transfer to unfamiliar problems, rather than merely making the model better at specific evaluations.

Second, Active Memory.

It is not the same thing as a 1-million-token context window.

The context window determines how much information a model can process in a single pass, whereas what Active Memory tries to solve is this: after days, weeks, or even months of interaction, can the model remember user preferences and proactively retrieve relevant information at the right time?

In the interview, host John Collison specifically used OpenClaw as an example, pointing out that existing Agents often have problems with long-term memory and context compression.

Wang responded that Meta designed both a memory system and model capabilities for this purpose. The model can not only use external memory, but is also trained to understand when memory is useful, how to perform context compression, and how to use different information sources to complete subsequent tasks.

However, there is currently no evidence that Meta has fully solved the long-term memory problem. Its MemoryBench internal tests use complex multi-turn conversations of about 64K to 100K tokens, which cannot be directly equated with sustained long-term use over months in real-world environments.

Third, Active Focus.

What it emphasizes is not simply lengthening the chain of thought, but enabling the Agent, during long-duration task execution, to continuously track goals, seek alternative approaches, and recover from errors.

Wang explained that these capabilities are closely related to task design in the model's post-training.

Meta uses supervised fine-tuning (SFT) to establish the behaviors the model should possess, and then uses reinforcement learning (RL) to increase the probability that the model stably exhibits these behaviors in real-world scenarios.

Among these, the quality of the reinforcement learning environment is especially important. Wang pointed out that if the environment incorrectly rewards certain unsafe shortcuts, the model may also learn behavior that does not meet expectations. Therefore, the team needs to carefully design training tasks, feedback mechanisms, and reward directions.

Meta also designed FocusBench specifically to evaluate an Agent's sustained execution capability when encountering obstacles, environmental changes, and task interruptions.

The technical direction truly worth noting here is this: Meta is no longer merely attaching an Agent workflow to a general-purpose language model, but optimizing from the pretraining and post-training stages for Agent capabilities such as long-term memory, sustained execution, and multimodal understanding.

Of course, Meta has not yet published the complete model architecture, parameter scale, or underlying memory algorithms, so outsiders still cannot reproduce these core technologies.

3D real-world tests expose shortcomings

Beyond the technical introduction, Meta also released a batch of practical application demos. The most intuitive among them is the 3D kitchen game. The model receives an image provided by the user and generates an interactive, playable three-dimensional kitchen scene.

Another demo showcased Agent tasks in health and lifestyle scenarios, as well as insurance claims assistance. The model needs to understand different materials, process information, and continuously advance multi-step work.

But after the release, developers quickly produced less flattering test results. X user Max Blade had Muse Wonder Full and Claude Opus 5.5 use the same prompt to each generate a AAA-grade simulation game in which one can drive a lawnmower while drinking beer and smoking a cigar.

The comparison video sparked considerable discussion.

Opus generated a relatively complete game scene with a cockpit, lawn, buildings, and so on, while Wonder Full's version was closer to a low-poly prototype, clearly inferior in materials and visual detail.

Some netizens joked that the two looked like the difference between a AAA game and a Roblox-style work.

But there is an easily overlooked detail: Meta has actually acknowledged this shortcoming. The Muse Wonder model card clearly states that visual and spatial reasoning, especially 3D tasks, remain a known limitation.

In response to community discussion, Alexandr Wang also acknowledged on X that the team did not specifically train the model for the 3D demo game, that in hindsight they should have strengthened this part, and suggested that developers use more real programming tasks to evaluate Wonder.

This also explains a seemingly contradictory result: Wonder Full can surpass Opus on long-horizon software engineering tests, yet may not be able to generate equally polished 3D games. The capabilities examined by the two tests are not entirely the same. The former emphasizes understanding and modifying complex engineering, while the latter also involves three-dimensional spatial understanding, scene design, and visual aesthetics.

However, this raises a question: if a general-purpose programming model needs targeted strengthening for different tasks, how will it perform in long-tail scenarios that are not sufficiently trained for in real applications, especially office scenarios?

Citrini, who had early testing access, said on X that Wonder strikes a good balance between capability and execution efficiency, and that its performance on financial tasks was also surprising. He is especially optimistic about Meta's opportunity in the small-business Agent market.

Shopify CEO Tobi Lütke also gave a positive assessment, saying that in his own tests, it was hard to find tasks where other top models failed and Wonder also could not complete them.

This divergence precisely shows that a model's capabilities cannot be fully summarized by a single composite score.

And for Wonder, which has been released for only one day, developer feedback is still an initial sample. Complete real-task reliability, average time consumption, and long-term performance still require more independent testing.

Costs rose by nearly 3 times

If Wonder's performance puts Meta back into the frontier model camp, then its cost makes this competition more realistic.

Currently, both models support a 1-million-token context window and input in text, image, video, and audio, with text as the main output.

The standard version Wonder is already available through the Meta Model API, Muse Code, and OpenRouter, priced at $4 per million input tokens, $0.2 for cached input, and $20 for output.

This price is on par with Claude Opus 5.5, twice the standard input and output price of GPT-6.1 Sol, and 60% lower than GPT-6 Astra.

But token-based pricing cannot directly reflect the actual cost of completing a task.

Artificial Analysis estimates that in comprehensive intelligence evaluations, Wonder consumes an average of about 7.69 million input tokens and 54,000 output tokens per task, involving multiple rounds of requests and tool calls, bringing the final task cost to $4.52.

The previous generation Muse Spark 1.3 cost only $1.60, GPT-6 Astra was $3.26, and GPT-6.1 Sol was only $0.72. In other words, Wonder's capabilities are clearly improved compared with Spark 1.3, but under this evaluation suite, task cost increased by about 183%, far exceeding the increase in the composite intelligence index.

Full is mainly provided through paid Muse Code plans, and developers can choose plans such as $50 per month for Power Usage or $200 for Maximum Usage. Previously, Max Blade's API-equivalent cost estimate in the 3D game test was about $4.11 for Wonder and about $99.43 for Opus, a difference of about 24 times.

But this does not mean that Wonder is 24 times cheaper for completing tasks of the same quality. On the one hand, the generated results are not the same; on the other hand, Full has no public API price, and its equivalent cost estimate cannot be conflated with actual subscription quota consumption.

Some developers also complained that Muse Code quotas were consumed too quickly in the early days after Wonder Full launched.

However, Alexandr Wang later publicly confirmed that this involved an anomaly in subscription quota calculation, and that the team had fixed the issue and reset quotas. Therefore, the abnormal early quota consumption should not be directly regarded as the normal usage cost of the model.

As for Muse, which targets ordinary consumers, Zuckerberg explicitly acknowledged in the interview that the Wonder model is larger in scale and has higher inference serving costs, so at this stage it is prioritized for paying users, while the free service continues to mainly use Spark.

He hopes that in the future, Wonder-level capabilities can be compressed into a lightweight form like Spark to serve hundreds of millions of consumers. AI industry signal channel: Frontier Technology. This content comes from the internet and the original link; the views represent only the author and do not represent Huxiu's position. If there are copyright issues, please contact hezuo@huxiu.com, and we will verify and handle them promptly. Everyone who is changing and wants to change the world is on the Huxiu APP[ ]]]