The full version of the Muse Wonder model scores 56 points on the Artificial Analysis Intelligence Index, tying for second place with several other models. However, the version available through the application programming interface (API), which companies can actually use, scores just 52 points, at a cost of $4 per million input tokens and $20 per million output tokens.
اضافة اعلان
The problem is that the index itself acknowledges that the overall score used by buyers may not reflect a model’s performance on their actual tasks.
The latest edition of the Artificial Analysis Intelligence Index shows very little difference among the leading models. Claude Opus 5.5 with Max settings ranks first with 58 points, followed by Claude Sonnet 5.5 with Max settings, Claude Opus 5.5 with Extra High settings, and Muse Wonder Full with Max settings, each scoring 56 points.
Among the 127 models evaluated, the gap between the top four is just two points. This narrow margin, however, creates a challenge for companies that rely on such rankings when deciding which AI solutions to purchase.
The headline-grabbing score of 56 applies to a version of the model available through specific subscriptions, while the version accessible to enterprise teams through the API scores only 52 points, at a cost of $4 per million input tokens and $20 per million output tokens.
More importantly, the index’s overall score may not be the most suitable measure of performance for the specific business tasks companies intend to run with the model, according to the index’s own methodology.
Narrow performance gaps make pricing and access more important
According to an analysis by Lina Park published in Forecast Mind, the small differences among leading models make pricing and the surrounding technical ecosystem among the most important factors distinguishing the options available to businesses.
The launch of Muse Wonder illustrates how different levels of access to a model can complicate enterprise purchasing decisions.
The full version of Muse Wonder, whose addition date is not specified by the index, is available only through monthly Muse Code subscriptions costing $50 for Power or $200 for Maximum. It is not available through the OpenRouter platform at all.
The landscape for open-weight models, meanwhile, looks very different.
According to the Artificial Analysis Intelligence Index, MiMo V2.6 Pro, the highest-ranked open-weight model, scores 46 points. Of the 127 models evaluated, 59 belong to the open-weight category.
This means the leading closed models have a 12-point advantage over the top open-weight model, leaving competition in the mid-tier performance category largely concentrated among closed models.
Do benchmarks reflect models’ real-world performance?
The Artificial Analysis Intelligence Index relies on a composite evaluation conducted by a single analytics organization rather than an independent, multi-organization audit.
The methodological notes published with the index explain that the scores represent model performance through the developer’s API or, when the developer does not provide a direct API, the median results reported by service providers.
But the most important caveat, which buyers may overlook, concerns how closely these results relate to different use cases.
According to the index’s guidance, while a model’s general intelligence tends to correlate with performance across a range of applications, specialized evaluations may be more relevant when assessing performance on particular tasks.
Version 4.3.2 of the index combines ten different evaluations to produce its overall score.
At the same time, the index publishes more specialized results, including AA-Bench’s version 1.1 evaluation of AI agents’ ability to perform knowledge work, and GDPval-AA version 2.1, which assesses practical tasks drawn from real-world work environments.
It also includes AutomationBench-AA, which evaluates automated workflows in software-as-a-service applications, and Enterprise Ops Gym-AA, which measures AI agents’ performance in corporate operational processes.
Relying solely on the overall score therefore means overlooking a range of indicators that may be more closely aligned with the tasks an organization wants to automate.
Enterprise use reveals a gap between benchmarks and reality
Actual usage data published on OpenRouter offers a glimpse into how Muse Wonder is being used.
According to the platform’s list, AI agent applications account for some of the largest volumes of tokens sent to the model.
Plato for Computer Use tops the list with approximately 11 billion tokens, followed by OMP with around 1.62 billion tokens and Enclave AI Penetration Testing with about 660 million tokens.
The list also includes Claude Code, with approximately 615 million tokens, and Kilo Code, with around 471 million tokens.
These figures suggest that the model is already being used for specific productive tasks, including computer control, software development and repeated code modifications.
However, benchmark results do not necessarily show how these systems fail during actual operations or who is responsible for detecting and addressing errors when they occur.
Human oversight reveals AI’s hidden costs
An audit published by The Brief Script, covering 150 enterprise AI deployments, provides figures illustrating the gap between technical capability and business outcomes.
The study concludes that technical implementation often advances faster than the development of mechanisms for measuring its business impact.
Of the 150 enterprise use cases that met the study’s criteria, 118 retained a clearly defined human role in approvals, exception handling, governance or active review.
By contrast, only 22 cases disclosed any information about costs or financial impact.
These findings suggest that the margin for error when deploying a mid-tier model does not disappear. Instead, the burden often shifts to employees responsible for reviewing and managing the system’s outputs.
Research into the cost of human review, discussed in an earlier report titled “Deployment Without a Clear View,” illustrates the scale of this burden when measured in practice.
According to the research cited in that report, the use of coding agents increased the volume of code produced by 30%, while the number of issues resolved remained unchanged.
At the same time, code review time rose by 49%, revision requests nearly doubled, and the number of employees involved in code review increased by 14%.
These findings indicate that producing more code does not necessarily translate into improved overall productivity, particularly when the output requires more scrutiny and correction.
Gartner expects more than 40% of AI agent projects to be canceled
According to research and consulting firm Gartner, more than 40% of AI agent projects will be canceled by the end of 2027.
The company attributes this forecast to rising costs, unclear business value and inadequate controls for managing risk.
The gap between expectations and reality becomes even more apparent when examining the views of senior executives.
According to figures cited in the report, 93% of leaders believe AI will eventually be able to manage core business functions, while only 7% have actually achieved that outcome.
The challenge with relying on mid-tier AI models lies in the distance between those two figures.
Overall benchmark scores are treated as if they certify a model’s readiness for enterprise use, while the costs of monitoring outputs and correcting errors are left to operational teams and insufficiently accounted for in purchasing decisions.
Five questions that matter more than a model’s ranking
For companies considering AI solutions, the most useful questions are not necessarily about comparing one model’s ranking with another’s, but about how the system will operate and how its behavior will be controlled.
The audit published by The Brief Script identifies five fundamental questions that organizations should answer before adopting AI agents in their workplaces.
These questions concern what information an agent can read, which tools it is allowed to select, which data or systems it can modify, which actions require prior human approval, and how it should handle situations in which its confidence falls or its actions fall outside approved policies.
Model rankings cannot directly answer any of these questions.
A score of 52 or 56 describes a model’s performance across ten specific evaluations, but it does not determine how that model will behave inside a real organization.
Ultimately, the nature of the work and the oversight mechanisms in place determine actual outcomes, while employees responsible for handling exceptions bear the consequences of errors and the gap between measured and real-world performance.
Source: Al-Ghad
