The End of Single-Vendor AI and the Rise of Routing
Multi-model routing is an AI cost optimization strategy where a central model gateway dynamically sends each request to the cheapest capable model, allowing production AI agents to scale without locking into a single provider or overpaying for frontier systems. Instead of building around one AI lab, leading engineering teams now treat models as interchangeable parts in a larger inference pipeline that can be swapped and routed as prices and capabilities change. Guillermo Rauch is blunt: picking one lab partner for everything "is a thing of the past," as companies stop burning tokens without regard for value and start insisting on vendor lock-in avoidance. The signal is clear—AI architecture is shifting from choosing "the best model" to choosing the smartest way to route tokens across many models.
Vercel and Coinbase Prove Multi-Model Routing Is a Cost Weapon
The most compelling evidence for multi-model routing comes from what builders are doing in production, not in demos. Vercel now routes more than a trillion tokens a day across millions of deployments and is "actively moving away from one-lab partnerships," using gateways to treat OpenAI, Anthropic, Gemini and others as plug-in components. Coinbase is making a similar architectural bet and has already cut its internal AI spend by nearly half while overall token usage kept rising, and engineers were not forced onto strict usage caps. In quotable terms: Coinbase reduced AI costs by nearly 50% using an internal LLM gateway that defaults to cheaper models and routes heavy tasks to frontier systems only when they are truly needed. Trying to pick a single best AI provider is, as their experience shows, a losing game.
How Model Gateways Turn Tokens Into a Managed Resource
Coinbase’s AI cost optimization playbook shows what intelligent model gateways can do when they are treated as a control plane rather than a thin proxy. First, an internal LLM gateway defaults engineers to lower-cost open-weight models like GLM 5.2 and Kimi 2.7, while still allowing frontier models when a job demands it. GLM 5.2, at roughly USD 1.40 (approx. RM6.44) per million input tokens and USD 4.40 (approx. RM20.24) per million output tokens, is three to six times cheaper per token than Anthropic’s Opus 4.8, at around USD 5 (approx. RM23.00) for input and USD 25 (approx. RM115.00) for output. Second, task-based routing ensures planning-heavy workloads go to stronger models, while simple execution flows stay on cheaper defaults. Third, by keeping chats bound to one model and leaning hard on caching, Coinbase raised cache hit rates from 5% to 60%, a 12x jump that removes enormous waste.
Gemini and the Economics of Multi-Vendor AI at Scale
Routing only makes sense if there are viable alternatives to the famous frontier labs, and that gap has now closed for many workloads. Rauch says he is seeing strong growth in Gemini because its models have "awesome price/performance characteristics" when companies scale up usage. At the same time, open-weight and Chinese models such as GLM-5.2 and Kimi 2.7 have become good enough for everyday engineering tasks while staying dramatically cheaper than flagship proprietary systems. Frontier models are now close in capability for routine work, but prices keep drifting apart, which makes multi-model routing the rational default instead of a niche optimization. This is why the model itself is turning into a commodity: when you can swap providers behind a gateway without touching application code, the real leverage sits in how you route and cache tokens, not whose logo is on the API.
Production AI Agents Need Infrastructure, Not Hero Models
The shift away from single-vendor stacks is driven by production reality, not hype. Rauch notes that last year was "all about prototyping" agents, but this year is about dealing with agents in production and their challenges—latency, reliability, cost and security. Coinbase now runs roughly 1,200 full-time production AI agents, normalized from compute hours, and at that scale Brian Armstrong argues human developers "have absolutely no business manually choosing which model to use"; the gateway must decide. Model gateways intercept every prompt, consider cache state, task complexity and real-time pricing, then route to the most cost-effective capable model. To make smart choices, teams need clear observability across providers: latency, uptime, token consumption and spend. The conclusion is hard to avoid—enterprise teams will win not by buying the most famous model, but by building infrastructure that turns multi-model routing into a core part of their production AI agents.
