Anthropic and OpenAI expanded enterprise model choice this week, with Claude Opus 5.5 matching the higher-end Fable 5.1 at lower operating cost and GPT-6 Sol and GPT-6 Luna splitting reasoning and high-volume work. Palo Alto Networks reported that no individual model detected more than 40% of vulnerabilities in complex environments in internal testing. The shift matters because IT teams must now decide which model fits each workload and when to switch, turning selection into architecture rather than one-time procurement.
How vendors tier models by cost and speed
Anthropic framed Claude Opus 5.5 as Fable-class performance without Fable-level running costs, while OpenAI positioned Sol for greater reasoning capability and Luna for faster throughput at scale. Both Sol and Luna carry lower API rates than their GPT-5.6 counterparts, with Luna aimed at higher-volume summarization, classification and extraction tasks. Salesforce illustrated the platform side by expanding support for multiple models, with its NVIDIA partnership bringing open models into Agentforce and Missionforce. Palo Alto Networks added a second proof point on Tuesday with a cybersecurity service that combines Claude Mythos, GPT-5.6-Cyber and open-weight models and assigns each security task to a chosen model.
The pattern separates prompts and workflow logic from the underlying models. An orchestration layer sits between applications and models and routes each request based on task, required capability, data restrictions and other application requirements. Because it handles requests across providers, the layer records which model processed each call and allows comparison on similar workloads. That separation lets a team replace a model without redesigning the application, as Jeet Pattanaik, founder and CTO of Glokal AI, described it. For agents the same idea applies at finer granularity, since planning, retrieving information, reasoning and taking action can each be routed to a different model.
Leading providers now compete through tiered portfolios built around price, performance and specialization rather than a single flagship model. Enterprise software vendors respond with platforms that accept models from several suppliers, reducing lock-in to one provider. Pattanaik noted that many portfolios emerge without central design, when separate teams choose providers for their own applications and governance is added afterward. Standby options also help when models are deprecated, terms change or access is interrupted. Palo Alto Networks testing supports the performance case: if each model finds vulnerabilities the others miss, combining them can lift coverage beyond the 40% ceiling observed for any single model.
What multi-model routing means for companies
For operating teams the direct effect is workload-level routing. A company can reserve a more capable model for complex reasoning or coding and send high-volume summarization, classification or extraction to a faster, lower-cost option such as Luna. OpenAI pricing for Sol and Luna against GPT-5.6 shows how large the gap can be across tiers. The approach also limits dependence on one supplier, since some workloads can move to a cheaper or stronger model without rebuilding every application. A small firm may apply this with two tiers and simple rules, while a large organization needs central routing, usage tracking and consistent access controls across departments and providers.
The constraint starts with data, not price. Pattanaik put data sensitivity first because it can rule out a model altogether, followed by reliability and latency, with capability and cost resolved by selecting the least expensive model that meets the task. Teams still must define which models can access specific data, monitor behavior across workloads and enforce security controls uniformly. Agent workflows add tracing risk: matching each step to a suited model helps quality, but an error from one model can become input for the next, which complicates diagnosis. Before committing, buyers should ask how routing decisions are logged, how performance is compared and how data boundaries are enforced.
A practical marker will be whether abstraction and measurement become standard practice. Pattanaik pointed to an abstraction layer that separates applications from models plus evals that track performance across workloads. If vendors and IT teams report routine cross-model comparisons, logged routing choices and clean swaps after pricing or availability changes, the multi-model stack has taken hold. Absence of those controls would signal that choice has grown faster than management.
