Coinbase and Vercel embrace multi-model routing to cut AI costs in half

Coinbase and Vercel are pursuing a strikingly similar architectural strategy: building systems that automatically route computational work across multiple AI models instead of locking in with a single vendor. This shift reflects a fundamental change in how companies are approaching AI infrastructure at scale.

Both organizations are moving away from the model-centric thinking that dominated early AI adoption. As capabilities among frontier models have converged and open-weight alternatives have improved dramatically, the economics increasingly favor flexibility over loyalty.

The trillion-token marketplace#

Vercel's CEO noted that his company now processes over a trillion tokens daily across millions of deployments, increasingly by spreading that load across competing providers rather than concentrating it with one lab. That shift represents a philosophical break from the partnership model that has dominated AI adoption so far.

Coinbase's experience demonstrates the financial impact. By redesigning its AI infrastructure with smart routing, the company managed to slash its AI spending roughly in half while simultaneously expanding token usage—without restricting engineers' access to more powerful models when they need them.

Three mechanisms for cost control#

The savings stem from three specific architectural levers:

• Default to cheaper models. Coinbase deliberately steers engineers toward more economical open-weight options like Z.ai's GLM 5.2 and Moonshot AI's Kimi 2.7 as the baseline. GLM 5.2 costs around $1.40 per million input tokens versus Anthropic's Opus 4.8 at approximately $5 per million input tokens—a 3-6x price gap. Despite the lower cost, these models perform competitively on code benchmarks, scoring 62.1 on SWE-bench Pro compared to other leading options. Because Coinbase self-hosts these models, sensitive code and query data remain entirely within the organization.

• Route by task type. Not every problem demands a frontier model's full reasoning power. Coinbase reserves expensive, advanced models for complex planning tasks while deploying cheaper alternatives for execution-focused work where both produce equivalent results. This pragmatic distinction eliminates unnecessary spending on premium capabilities for routine operations.

• Aggressive context caching. By keeping conversations locked to the same model while cached context remains valid, Coinbase increased its cache hit rate from just 5% to 60%—a 12x improvement that dramatically reduces redundant processing costs.