On October 12, 2026, Meta announced a new AI model, Muse Wonder, along with a higher-tier version, Muse Wonder Full. On an independent evaluator's composite index, Full scored on par with Claude Sonnet 5.5 and ahead of GPT-6 Astra, making Meta a serious contender in the race for frontier AI models. However, developers can access only the standard Muse Wonder through the API. Full is limited to paid plans of the coding assistant Muse Code. Looking closely at the performance gap between the two models, and at their strengths and weaknesses against rival models, shows how this announcement may affect the choice of AI model.
Muse Wonder Full scores 56 on the composite index; the API version scores 52
On the Intelligence Index from Artificial Analysis (AA), an independent evaluator of AI model performance, Muse Wonder Full scored 56 and the standard Muse Wonder scored 52. The previous model, Muse Spark 1.3, evaluated at the same max setting, scored 48, so Meta has steadily raised its performance.
The index combines the results of 10 evaluations covering knowledge, reasoning, coding, professional tasks and more. Comparing the major models listed in the AA model list as of October 13, 2026 shows that Full and the standard version occupy quite different competitive positions.
| Model | Reasoning setting used in evaluation | AA composite index | AA cost per evaluation task |
|---|---|---|---|
| Claude Opus 5.5 | max with fallback | 58 | $5.98 |
| Claude Sonnet 5.5 | max with fallback | 56 | $5.46 |
| Muse Wonder Full | max | 56 | Not listed |
| Claude Fable 5.1 | max with fallback | 53 | $7.63 |
| GPT-6 Astra | max | 53 | $3.26 |
| Gemini 4 Argon | high | 53 | $1.99 |
| Muse Wonder | max | 52 | $4.52 |
| GPT-6.1 Sol | max | 52 | $0.72 |
| Muse Spark 1.3 | max | 48 | $1.60 |
A higher index means better performance. However, Claude's "with fallback" is an evaluation condition that includes switching to an alternative model, so this is not a comparison of all models run in the same environment. Nor does each company's "max" necessarily mean the same amount of compute. As for cost, AA's figure is a weighted average of per-task costs across evaluations, which differs from the actual cost of obtaining one successful result in real work. Detailed definitions are available on AA's Muse Wonder Full evaluation page.
Full's composite score of 56 falls short of Claude Opus 5.5's 58 but matches Claude Sonnet 5.5.
The standard Muse Wonder, by contrast, scored 52: one point below GPT-6 Astra, Claude Fable 5.1 and Gemini 4 Argon, and tied with GPT-6.1 Sol.
In other words, just because Full matches top-tier models, that assessment cannot be applied as-is to the standard version available through the API.
Both models support context windows of up to 1 million tokens. They accept images, audio and video as well as text, but output is limited to text.
According to Meta's model card, integration into the Muse app is planned, with availability rolling out to paid subscribers first. What the model itself can do should be distinguished from what is actually available in each service.
Strong in long-running software development, but weaker in some areas
The Muse Wonder series showed particularly strong performance in long-horizon development work: investigating existing software and repeatedly modifying and testing code.
In comparison results published by Meta, even the standard Muse Wonder beats the Claude and GPT models it was compared with on both DeepSWE 1.1 and SWE-Marathon.
| Development benchmark | Wonder Full | Wonder | Fable 5.1 | GPT-6 Astra | Opus 5.5 |
|---|---|---|---|---|---|
| DeepSWE 1.1 | 79.0 | 78.8 | 67.4 | 74.1 | 74.2 |
| SWE-Marathon | 57.5* | 53.1 | 45.6 | 42.5 | 46.3 |
| Terminal-Bench 4.0-AA | 57.1 | 52.0 | 52.0 | 59.1 | 59.6 |
| ProgramBench | 59.0 | 50.0 | 55.5 | 52.5 | 70.5 |
Figures are based on Meta's comparison table at launch and Section 3 of the model card. They show pass rates or accuracy set by each benchmark, as percentages; higher is better. Evaluations were generally run at the max setting, but Full's SWE-Marathon result, marked with *, used the xhigh setting. Third-party measurements, figures published by each developer and Meta's own measurements are mixed, and the execution environment differs by test.
DeepSWE 1.1 evaluates development results through functional and regression tests on 113 tasks across 91 repositories.
Here Full scored 79.0 and the standard version 78.8, a gap of only 0.2 points. According to Meta's model card, third-party measurements by Datacurve were used for both Muse Wonder models and Claude Opus 5.5. For GPT-6 Astra and Claude Fable 5.1, the figures published by their respective developers were used.
On SWE-Marathon, which targets development projects spanning a longer period, Full scored 57.5 and the standard version 53.1.
However, these are results evaluated by Abundant AI, with GPT-6 Astra run on Codex and Claude Fable 5.1 on Claude Code. Only Full used the xhigh reasoning setting.
The results therefore suggest that the Muse Wonder series is strong at long-horizon development work, but they are not a like-for-like comparison of pure model performance.
On Terminal-Bench 4.0-AA, which tests complex work in a terminal environment, Claude Opus 5.5 scored 59.6 and GPT-6 Astra 59.1, both ahead of Full's 57.1.
On ProgramBench, which tests reproducing programs in a clean-room environment, Opus 5.5 leads by a wide margin at 70.5. Full scored 59.0 and the standard version 50.0.
In ProgramBench, all models used the same mini-swe-agent, and the metric is the share of attempts that passed at least 95% of tests.
These results show that using Muse Wonder Full will not improve every kind of development work.
The gap with the standard version is only 0.2 points on DeepSWE but widens to 9.0 points on ProgramBench.
The standard version available through the API also has strengths in long-horizon development, but models should not be chosen on the vague verdict that they are "strong at coding."
Whether the task is fixing existing code, operating a complex development environment or reproducing an entire program, the best model changes depending on the specific work.
What a 13.1% legal score and an 85.5 on a memory test mean
Muse Wonder Full also performs well in legal work and business automation. However, other companies' models came out ahead on specialized health questions and on operating a computer screen.
Extracting the evaluations related to professional work and personal AI agents from Meta's published comparison table gives the following.
| Evaluation | Wonder Full | Wonder | Fable 5.1 | GPT-6 Astra | Opus 5.5 |
|---|---|---|---|---|---|
| Harvey LAB-AA v1.1 (legal) | 13.1 | 10.0 | 6.4 | 8.6 | 4.2 |
| Vals Finance Agent V2 (financial research) | 59.1 | 56.6 | 58.9 | 53.5 | 58.6 |
| HealthBench Pro (specialized health questions) | 62.3 | 60.5 | 62.1 | 65.3 | 65.6 |
| AutomationBench-AA (business automation) | 45.4 | 39.1 | 32.1 | 41.6 | 41.2 |
| OSWorld V2.1 (screen operation, partial credit) | 78.4 | 72.6 | 80.7 | 75.8 | 81.8 |
| MemoryBench (recall of conversation content) | 85.5 | 85.6 | 79.0 | 81.4 | 81.5 |
All were evaluated at the max setting. Figures are the published scores for each benchmark; higher is better. However, evaluation methods differ, for example full pass rates versus average partial credit, so numbers cannot be compared directly across rows. Details on who ran each evaluation and how it was scored are in Meta's model card.
One figure deserves particular attention: Full's 13.1 in legal work.
Harvey LAB-AA v1.1 measures the share of deliverables, produced from case materials, that meet all scoring criteria and also pass a check for content not grounded in fact.
Full beat the models it was compared with, but only 13.1% of tasks met every condition.
Improved ability to help draft professional documents is a completely separate matter from being able to hand work over without human review.
AutomationBench-AA, which evaluates business automation, shows 45.4 for Full, again the share of tasks in which work executed in simulated applications met all conditions.
By contrast, OSWorld V2.1's 78.4 is an average of partial credit. Judged only on whether a task was fully completed, Full's score is 39.4.
Even for the same kind of AI agent evaluation, the meaning of a number changes substantially depending on whether it rewards getting partway through a task or finishing it. The results shown here are based on the scoring criteria at the time of Meta's announcement.
One feature Meta emphasizes for its personal AI agent is "Active Memory."
This refers to the ability to retain important information from a long conversation and retrieve it accurately when needed.
In Meta's MemoryBench, the standard Muse Wonder scored 85.6 and Full 85.5, ahead of the Claude and GPT models compared.
However, MemoryBench is an internal Meta evaluation. Irrelevant topics and information updated midway are inserted into long conversations of 64,000 to 100,000 tokens, and another AI scores, under set criteria, whether the model correctly recalls the necessary facts.
The result shows a high ability to extract needed information from long conversations. It does not prove that the model can remember a user's information accurately over months, or that its save and delete functions are reliable.
The 1-million-token context length and the memory function in continued use of a service should also be considered separately.
Another feature, "Active Focus," refers to the ability to try a different approach when stuck during work and to keep going while performing the necessary checks.
Meta explains that, in addition to training on long-horizon tasks, adopting a base model more capable than Muse Spark spread the improvement to tasks not specifically targeted in additional training.
According to Meta's research blog, the "Prometheus" data center, with gigawatt-class power capacity, was used for pretraining and reinforcement learning.
However, the model's parameter count has not been disclosed, so it is impossible to judge whether it is larger or smaller than competing models.
API pricing is 40% of GPT-6 Astra's, but actual processing costs differ
Muse Wonder's API pricing is set at $4 per million input tokens and $20 per million output tokens.
The standard input and output rates match Claude Opus 5.5 and equal 40% of GPT-6 Astra's. Compared with Muse Spark 1.3, which was known for low prices, it is a substantial increase.
Comparing Meta's official price list as of October 13, 2026 with competitor API prices listed by AA gives the following.
| Model | Input, per 1M tokens | Output, per 1M tokens | Cached input, per 1M tokens |
|---|---|---|---|
| Muse Spark 1.3 | $1.25 | $4.25 | $0.15 |
| Muse Wonder | $4 | $20 | $0.20 |
| Claude Opus 5.5 | $4 | $20 | $0.20 |
| GPT-6 Astra | $10 | $50 | $1 |
Meta's prices are for Standard mode. The cached input price applies to reused input and does not include all costs such as writing to the cache. Full's pay-as-you-go API pricing has not been published, so it is not included in the table.
The key point is that a model with lower per-token pricing does not necessarily process real work more cheaply.
In AA's evaluation as of October 13, 2026, Muse Wonder's cost per task was $4.52. That is about 1.39 times GPT-6 Astra's $3.26 and about 2.83 times Muse Spark 1.3's $1.60.
These ratios were calculated from the costs listed in AA's evaluation list by dividing $4.52 by each comparison model's cost, rounded to two decimal places. All were evaluated at the max reasoning setting.
AA divides costs such as the input and output tokens actually used, and caching, by the number of evaluation tasks, then takes a weighted average according to the weight of each evaluation within the composite index. The detailed calculation method is explained on AA's model evaluation page.
This metric does not represent the cost per successful piece of work. Even so, it is clear that API per-token pricing alone cannot determine actual usage costs.
Developers choosing an API also cannot overlook that GPT-6.1 Sol, which has the same composite index of 52, costs $0.72 per evaluation task.
However, models with the same composite index will not necessarily deliver the same quality on individual tasks. Response speed and the number of tokens consumed also differ.
If Muse Wonder is used for long-horizon software development, it is necessary to measure the total cost of reaching the final result, in addition to how much it raises task success rates and reduces the effort of corrections.
Muse Wonder Full's pricing, meanwhile, works differently from the API.
Muse Code, which supports Full, offers "Power Usage" at $50 per month and "Maximum Usage" at $200 per month. Both are subscriptions with usage limits and cannot be compared directly with API per-token prices. Availability conditions are explained on the official Muse Wonder page.
Handling of input data also differs depending on the service and the type of contract.
According to Meta's explanation of pricing and usage modes, data sent through the API's Standard mode is not used to train models. The Wonder API also has no Contributor mode, which offers lower prices in exchange for contributing to training.
In Muse Code subscriptions, by contrast, Contributor mode is enabled by default and can be switched off in settings.
When handling corporate source code or confidential documents, these data-handling settings are an important factor in the decision, alongside whether Full is available.
Meta returns to the frontier AI race; real-world performance will matter in choosing a model
The Muse Wonder series shows that Meta has strengthened its presence in the race for frontier AI models.
In long-horizon software development in particular, even the API-accessible standard version scored highly. On ProgramBench, which evaluates program reproduction, the gap with Full widens, so the value of choosing the higher-tier version depends on the use case.
In legal work, it posted a higher full-pass rate than competing models, but that rate is only 13.1%. And on specialized health questions and screen operation, other companies' models still hold the advantage in some evaluations.
In short, Meta has added a strong option for handling professional work and long-horizon development, but it has not built a model that beats rivals in every use case.
Meta's model card also explains that performance varies with the execution environment, the tools used and how instructions are given. It also states that evaluations at the same level were not conducted for every language used in training.
When using the model for Japanese-language internal documents or real development projects, it is worth checking not only how generated text and code look, but also whether it responds appropriately when conditions change midway and whether the final deliverable meets requirements.
Whether embedding the standard Muse Wonder through the API or using Full in Muse Code, what matters is not the published benchmark rankings but how far the model can improve the balance of performance and cost in actual work.
If the strengths shown here are reproduced in practice, the Muse Wonder series could become a strong option as an AI to handle everyday software development and complex business tasks.