Multi-Model Routing: The New Default for Enterprise AI
Multi-model routing strategy is an AI deployment approach where enterprise systems treat language models as interchangeable components, dynamically sending each task to the most suitable model based on price, performance, latency, and security requirements instead of relying on a single provider relationship or manual model selection.
That shift is no longer theoretical. Vercel CEO Guillermo Rauch and Coinbase CEO Brian Armstrong now share the same opinionated bet: single-lab partnerships are obsolete, and intelligent, vendor-agnostic AI deployment is the only way to scale agents economically. Vercel routes more than a trillion tokens a day across millions of deployments and is actively moving away from one-lab partnerships. Coinbase, meanwhile, cut internal AI spend by nearly half while token usage kept growing and engineers faced no usage caps. This is not AI agent cost optimization as a side project; it is a strategic rewiring of enterprise AI infrastructure. Trying to pick the single best AI provider, as one source bluntly puts it, is a losing game.
Coinbase: 1,200 AI Agents, 50% Cheaper
Coinbase’s architecture shows what serious AI agent cost optimization looks like at production scale. The company now runs roughly 1,200 full-time AI agents, a figure derived by normalizing compute hours to a 40–60 hour workweek. At that scale, Armstrong argues that human developers have no business manually choosing models; the infrastructure must automate those decisions.
The core of their multi-model routing strategy is an internal LLM gateway that defaults engineers to lower-cost open-weight models, specifically Z.ai’s GLM 5.2 and Moonshot AI’s Kimi 2.7. GLM 5.2 is priced at roughly USD 1.40 (approx. RM6.40) per million input tokens and USD 4.40 (approx. RM20.20) per million output tokens, versus Anthropic’s Opus 4.8 at around USD 5 (approx. RM22.80) for input and USD 25 (approx. RM114.00) for output. That is a three- to six-times cost reduction per token while still posting a 62.1 score on SWE-bench Pro compared to GPT-5.5’s 58.6. The result: Coinbase cut internal AI spend by nearly half without slowing token usage growth.
Two more levers make this work. First, task-based routing: frontier models handle complex planning, while cheaper models take routine execution work where they perform equally well. Second, aggressive caching: by sticking with the same model while context stays valid, Coinbase pushed cache hit rates from 5% to 60%, a 12x jump that drives costs down further. This is token usage scaling by design, not by accident.
Vercel: AI Gateway as the New Control Plane
If Coinbase proves the financial upside, Vercel shows the infrastructure pattern every serious engineering team will copy. Its cloud platform handles 6 million deployments daily, with roughly half triggered by coding agents, and more than 1 trillion tokens passing through its AI gateway each day. Rauch’s stance is blunt: models are now interchangeable parts inside a larger inference pipeline, and one-lab partnerships are outdated.
This is enterprise AI infrastructure recast as a control plane. The gateway intercepts every prompt and decides, in milliseconds, whether the job needs an expensive frontier model or a cheaper alternative based on cache state, task complexity, and real-time pricing. As one source describes it, the competitive advantage shifts away from the model and toward the infrastructure that decides which one to use. Vercel’s Eve framework lets companies define agent instructions and skills in natural language, while its Sandbox limits what data agents can access or export to keep sensitive information from leaking into model training. A sales representative using an internal agent to spot fast-growing accounts shows how AI adoption is often constrained more by data access than by analytical ability.
Crucially, Vercel’s customers are not staying loyal to any single lab. Rauch reports growing use of Gemini, DeepSeek, and GLM-5.2 alongside OpenAI and Anthropic as enterprises tune for price-to-performance rather than exclusivity. That is vendor-agnostic AI deployment in practice.
Why Single-Vendor AI Is Losing and What Enterprises Should Do Next
The timing of this shift is not accidental. Frontier models have converged in capability for day-to-day engineering work, open-weight alternatives have improved sharply, and the price gap keeps widening. When GLM 5.2 can match or beat GPT-5.5 on key coding benchmarks while costing three to six times less per token, locking into a single provider is less strategy and more complacency. Meanwhile, enterprises now expect token usage scaling into the trillions without runaway bills; Coinbase proves it is possible to cut spend by half even as usage rises.
The deeper change is philosophical. Both Vercel and Coinbase assume today’s “best” model will not stay on top for long. In that world, loyalty is risk. Trying to pick the single best AI provider is, in the words of one source, a losing game. The winning bet is vendor-agnostic AI deployment built on gateways that abstract providers, automate routing, and keep observability front and center. Teams need clear metrics on latency, uptime, token consumption, and cost across all providers, because without that data, it is impossible to know if routing choices are helping or hurting.
The takeaway is unapologetically opinionated: if your AI roadmap still hinges on a single-vendor deal, you are building technical debt into your cost structure. Follow the playbook emerging from Vercel and Coinbase instead: deploy an internal model gateway, default to cheaper models and use frontier models sparingly, route tasks by complexity, cache aggressively, and treat evaluation as an ongoing process tuned to your own workloads. Multi-model routing is not a fancy add-on. It is fast becoming the minimum viable architecture for serious enterprise AI.
