Routing Moved From the Architecture Diagram to the Product
I wrote "Decoupling AI Agent Logic from Model Providers: The Dispatcher Pattern" as an engineering argument. Model providers change often, agent logic changes slowly, and a routing layer between them turns migrations into configuration changes. The audience was a team deciding how to structure its own system.
Two items from early October suggest the pattern has outgrown that framing. Routing has become part of what companies sell.
The first came from xAI. Elon Musk announced that Grok Bot will route user tasks to third-party models: Anthropic's Claude Opus 5.5 for heavy reasoning and coding, Midjourney for images, and Suno for music. All-In called it going headless, and the panel debated what it signals. One reading says xAI lacks confidence in Grok across all task types. The other says xAI is making a platform bet that the routing layer matters more than winning every model comparison. The second reading got less airtime and strikes me as the more interesting one, since it resembles how marketplaces work. The company that owns the customer relationship and the routing decision captures value regardless of which supplier executes the task.
That is a notable inversion for the dispatcher pattern. In my original framing, the dispatcher sits inside an organization's architecture, invisible to end users, and its purpose is cost and resilience. In Grok's case the dispatcher is the product, and end users buy an experience whose underlying model they never see. One of Anthropic's most capable generally available models sits inside a competitor's consumer app, largely invisible to its users.
The second item concerns the buyers. The Information reported that Meta's internal Claude Code usage fell from about 60,000 users to about 30,000, and that Microsoft is also steering staff toward in-house models. I only have the report's account, behind a paywall, and the counts are unconfirmed, so I hold them loosely. The direction is still plausible and consistent with the pricing pressure across the tier stack: the largest buyers have the volume and the engineering depth to build their own alternatives and shift traffic when the economics change.
Taken together, the two items show routing happening at both ends of the market. Sellers route among suppliers to build a product. Buyers route among internal and external options to manage cost. Neither group commits wholly to one model, and each has built or bought the layer that lets it change its mind.
What changes the story for ordinary enterprises is what the routing layer accumulates. A dispatcher that has handled millions of requests knows which task types succeed on which models, which prompts need adapting, how often each provider fails, and what each path costs. The routing table in my original post is described as a configuration artifact, and in practice it becomes a record of institutional knowledge about model behavior. The organization that owns that record can switch suppliers without starting from zero.
Mixture of Experts made a related observation about Sonnet 5.5 and the mid-tier squeeze: routing decisions, which model for which task, are becoming work every enterprise team has to own. When models approach each other in capability and converge on price, as the late-September launches did, selection moves from choosing a winner to maintaining a mapping. The mapping needs evaluation data to stay accurate, which ties back to the evaluation harness the original post depends on.
I would amend the original post in one respect. It treated the dispatcher mainly as insurance against change. The October evidence suggests it also works as an asset: the layer where cost control, quality evidence, and supplier leverage accumulate. An organization with a mature routing layer holds a bargaining position in every renewal conversation, since a credible alternative is a configuration change away.
xAI, Meta, and Microsoft are already making the same bet: they split work across models, hold the routing decision themselves, and treat any single model as replaceable. A pattern that originated as an engineering convenience has become a commercial structure.
Sources: All-In Ep. 292, "Is Claude Conscious? Pope Rejects, Model Welfare Movement, OpenAI's Math Backlash, France Riots" (Oct 10, 2026); The Information, "Meta, Microsoft work to wean staff off Anthropic's Claude" (Oct 2, 2026); Mixture of Experts Ep. 127 (Oct 3, 2026). Usage counts are as reported by The Information and were not independently confirmed.