Anthropic, OpenAI launches show shift toward multi-model enterprise AI

Multi-model AI moves into enterprise software

The shift toward multi-model AI is also shaping enterprise software, as vendors build platforms that support models from multiple providers. These platforms give enterprises more flexibility to choose models for different applications without tying their workloads to a single provider.

For example, Salesforce has been expanding its AI platform to support multiple models and enterprise-specific AI capabilities. At the same time, its partnership with Nvidia brings open models into its Agentforce and Missionforce platforms.

Palo Alto Networks offered another example on Tuesday, introducing an AI cybersecurity service that combines Anthropic's Claude Mythos, OpenAI's GPT-5.6-Cyber and open-weight models. The system determines which model to use for each security task rather than relying on a single model for the entire workload, according to the company.

In internal testing, Palo Alto Networks found that no individual model detected more than 40% of vulnerabilities in complex environments. That suggests a single model could miss many vulnerabilities, while combining models could improve coverage if each identifies vulnerabilities that the others miss.

Beyond model performance, using multiple providers can also give enterprises alternatives when models are deprecated, terms change or access is interrupted, Pattanaik said. 

From choosing a model to managing models

Using multiple models can give enterprises more flexibility, but it also creates a new management challenge. IT teams must decide not only which model fits a particular workload, but also how to manage those models across the broader AI environment.

Model selection affects more than application performance. IT teams also need to consider cost, latency, data access, security controls and the type of work the model needs to perform.

"Data sensitivity should be the first consideration because it can rule out a model altogether, followed by reliability and latency," Pattanaik said. Capability and cost then come down to finding the least expensive model that can meet the requirements of a particular workload, he added.

A company might use a more capable model for complex reasoning or coding while directing high-volume summarization, classification or extraction tasks to a faster, lower-cost model. OpenAI's new pricing for Sol and Luna underscores the potential cost differences: both models carry lower API rates than their GPT-5.6 counterparts, with Luna designed for higher-volume workloads.

A multi-model approach can also reduce an organization's dependence on any single provider. If a new model offers better performance or lower costs, an enterprise can potentially route some workloads to it without replacing every AI application in its environment.

But that flexibility introduces another layer of infrastructure to manage.

IT teams need to determine which models can access specific data, monitor how models perform on different workloads and apply consistent security and governance controls across providers.

"Organizations can replace a model without redesigning the application if they keep prompts and workflow logic separate from the underlying models," Pattanaik said.

That setup puts an orchestration layer between applications and models, determining which model should manage each request based on factors such as the task, required capability, data restrictions or other application requirements.

The orchestration layer can also help enterprises track model performance across different tasks. Because it manages requests across models, IT teams can see which model handled each request and compare how they perform on similar workloads, Pattanaik said. 

AI agents add another layer to model routing

Model routing becomes more complex when AI agents manage multiple steps within a single workflow. An agent can break a workflow into tasks such as planning, retrieving information, reasoning and taking action, with routing occurring at each step.

Using several models can help match each task with a model suited to it, Pattanaik said, but it can also make failures harder to trace. If one model produces an incorrect output that another model uses as input, IT teams may need to determine which step caused the problem.

This makes model routing part of an AI application's architecture, rather than simply a way to optimize performance or cost. It also underscores the importance of separating applications from the models they use. As providers change model performance, pricing and availability, organizations can switch the model executing a particular task without necessarily changing the application itself.

This flexibility can become increasingly important as enterprises build AI applications that may outlive individual models or providers. Instead of standardizing on a single model, IT teams might need systems that can evaluate, route and govern models as they evolve.

An abstraction layer can separate an application from the specific models it uses, while evaluations can help IT teams measure model performance across workloads.

"The important part is building an abstraction layer that separates the application from the models and using evals to measure how those models perform," Pattanaik said.

Kinza Yasar covers AI and emerging technology for TechTarget, with a focus on ethics, enterprise adoption, governance and business strategy. Before moving into journalism, she worked in IT and network support roles, giving her a systems-level perspective on how enterprise technologies are built, deployed and managed.