Multi-Model AI Routing: From Hype Spend to Production Discipline
Multi-model AI routing is an infrastructure strategy where enterprises route workloads across several language models through a central gateway, matching each task to the model with the best mix of price, performance, and policy constraints instead of committing to a single AI vendor for all use cases. The headline story is simple: the era of picking one AI lab and building everything on it is ending, and companies that cling to single-vendor stacks will overpay for mediocre outcomes. Vercel’s CEO says companies are no longer relying on one lab for all their needs, because every part of the AI stack is now “plug and play”. That shift is not theoretical; it is a profit-and-loss decision. Money poured into AI is no longer excused as experimental spend when it fails to translate into value shipped.
Vercel and Coinbase: Single-Lab Loyalty Is Officially Dead
Two very different companies, Vercel and Coinbase, have converged on the same architecture: model gateways that can route work across multiple providers. That is not a minor tweak; it is a rejection of the old “pick a lab and pray it stays ahead” mindset. Guillermo Rauch says Vercel now routes more than a trillion tokens per day across millions of deployments while actively moving away from one-lab partnerships. Coinbase CEO Brian Armstrong has taken the same route, cutting his company’s internal AI spend by nearly half while token usage continued to grow and without capping engineers’ access. One quotable lesson emerges: trying to pick the single best AI provider is a losing game. In other words, loyalty is no longer a strategy; routing is.
Gemini, Open-Weight Models, and the Price-to-Performance Arms Race
The reason multi-model AI routing works is that the model landscape has flattened for many day-to-day tasks while prices have diverged sharply. Frontier models have become much closer in capability for everyday engineering work, open-weight alternatives have improved, and the price gap keeps widening. Rauch says he is seeing strong growth in Gemini because its models have “awesome price/performance characteristics” when scaling up. Coinbase makes that calculus explicit: its internal gateway defaults to lower-cost open-weight models such as GLM 5.2 and Kimi 2.7, reserving expensive frontier models only when a job demands them. GLM 5.2 is priced at about USD 1.40 (approx. RM6.40) per million input tokens and USD 4.40 (approx. RM20.00) per million output tokens, compared to Anthropic’s Opus 4.8 at around USD 5 (approx. RM23.00) for input and USD 25 (approx. RM115.00) for output, a three- to six-times reduction per token.
Gateways, Smart Routing, and Vendor Lock-In Avoidance
The core of this new architecture is an internal model gateway strategy. Coinbase runs an LLM gateway that acts like a control plane: it intercepts every prompt and makes a split-second decision about whether the workload needs an expensive frontier model or a cheaper alternative. Armstrong argues human developers should not be manually choosing models at this scale, especially when they already operate the equivalent of around 1,200 full-time AI agents in compute hours. Task-based routing pushes complex planning to stronger models while sending simple execution work to cheaper ones, because there is “zero reason to pay top dollar” when cheaper models perform just as well. This is vendor lock-in avoidance in code. Rauch explicitly compares today’s shift to the move from single-cloud to multi-cloud: companies partner with different AI labs for different tasks to reduce reliance on one vendor and optimize costs.
AI Cost Optimization as an Operating Principle, Not a Phase
Enterprises have discovered that telling teams to burn tokens without constraint does not guarantee customer value; the free-spending prototyping phase is over. Coinbase’s experience is instructive: by combining cheaper default models, task-based routing, and aggressive caching, the company halved its AI bill while total AI token usage kept growing. Cache hit rates jumped from 5% to 60%, delivering a twelve-fold improvement in reuse and making routing decisions even more important. Vercel’s perspective reinforces the pattern: the model is now one interchangeable component in a larger pipeline, and single-lab partnerships are obsolete. Gemini’s growing role, alongside Chinese models such as DeepSeek and GLM-5.2, underlines a broader recognition that no single provider dominates every workload and price point. The conclusion is blunt: AI cost optimization is now a structural requirement, and multi-model AI routing is how serious enterprises meet it at scale.
