The race to build the best artificial intelligence model may already be over, and the real competition is shifting to something far less glamorous: the software layer that decides which model to use for each task. According to technology analysts Ben Bajarin and Jay Goldberg, who unpacked this thesis on a recent episode of The Circuit podcast, frontier models like GPT-6, Claude 5, and Fable are becoming so closely matched in performance that capital investment, not innovation, now governs which model leads on any given benchmark.

The implications are profound for enterprises, startups, and anyone building AI products. If models are commoditizing, then the durable competitive advantage will accrue not to OpenAI, Anthropic, or Meta, but to whoever controls the orchestration layer: the intelligent router that sends simple tasks to free open-source models, complex reasoning to frontier models, and crucially, remembers everything about the user across all those switches.

What Does "Model Commoditization" Actually Mean?

Over the week of July 8-15, 2026, a wave of model releases hit the market: OpenAI's GPT-5 and GPT-6, Meta's next-generation models, a competitive Grok variant, and continued access to Anthropic's Fable, which was originally scheduled to be shut down but has now postponed decommissioning three times. The sheer volume of capable models arriving simultaneously signals a market where performance gaps are real but narrowing rapidly.

Creative Strategies, Bajarin's firm, published a benchmarking tool at csbench.com that evaluates models on practical knowledge work tasks: creating PowerPoint presentations, Excel sheets, and Word documents from corporate data. The key finding is not just about output quality, but about efficiency. The real variance lies in how many tokens, or computational units, it takes to reach a polished result.

For knowledge work, the average user sees little difference between models on first glance. But when you measure tokens-to-outcome, the picture shifts. Models are already training for efficiency: they spawn fewer unnecessary sub-agents and use fewer tokens per correct answer. This is the new frontier of competition, and it has profound implications for enterprise adoption, where token budgets are finite.

How to Think About the Orchestration Layer Advantage?

• Task Routing: The orchestration layer directs simple work to free or cheap open-source models like DeepSeek or GLM, reserving expensive frontier models for complex reasoning tasks that genuinely require their capabilities.
• User Memory Retention: All files, sessions, and preferences follow the user across model switches, creating a form of lock-in that is difficult to replicate. If your files and conversation history live in one system, switching to a competitor becomes friction-filled.
• Data Ownership: The orchestration layer, not the model provider, owns the user's data memory. This is the ultimate franchise: knowing everything about a user's work patterns, preferences, and history.

"I use Claude because all my files are there. ChatGPT has access now, but it doesn't know as much about me. That's a form of lock-in," explained Jay Goldberg, attorney and analyst at D2D Advisory.

Jay Goldberg, Attorney and Analyst at D2D Advisory

Goldberg made a sharp prediction: the enduring winner in AI will not be the company with the best model. Models will remain closely competitive, governed largely by capital. The real advantage will accrue to whoever builds the orchestration layer that routes tasks to the optimal model for the job while retaining the user's memory across model switches.

The orchestration layer is easy to build from a software perspective, a trivial technical problem. But it is hard to adopt because it requires model-agnosticism from the frontier labs themselves, which they may resist. If a third party builds it and wins the data-memory franchise, that company could dominate the AI application market.

Why Enterprises Will Force This Change?

Bajarin noted that enterprises will not leave model selection to employees. Shadow IT will resurface if the internal tool is hostile or slow. The average employee just wants the task done; they will not switch models manually. This creates pressure for a unified interface that handles model selection invisibly, in the background.

Goldberg experienced this firsthand at a large mutual fund company's AI system. When the internal tool was cumbersome, employees found workarounds and used competing services on their own time. The lesson is clear: if the orchestration layer is not seamless and fast, users will abandon it for something better, regardless of what management mandates.

The broader implication is that NVIDIA's dominance in AI chips, and the race to build the best large language model (LLM), may matter less than the race to own the user interface and data layer above those models. The company that controls how users interact with AI, and that retains their data across model switches, may ultimately prove more valuable than the company that trained the model itself.