Coinbase Cut Its AI Bill in Half by Routing Around Anthropic and OpenAI. The Model Just Became the Least Interesting Part of the Stack.

Coinbase just proved that the most expensive line item in enterprise AI — the foundation model itself — is now optional. The company halved its AI infrastructure bill by routing a large share of its workloads to two Chinese open-weight models, Zhipu's GLM 5.2 and Moonshot's Kimi 2.7, while token usage kept climbing.

The mechanism matters more than the headline savings. Coinbase didn't ask engineers to pick a cheaper model and accept worse output. It built a routing layer that inspects every prompt before it's sent anywhere, matches task complexity, cache status, and per-token pricing against a menu of models, and picks the cheapest one capable of doing the job. CEO Brian Armstrong's framing was blunt: humans shouldn't be choosing models — that's now a machine decision.

The price gap explains the incentive. GLM 5.2 runs $1.40 per million input tokens and $4.40 per million output tokens. Anthropic's Opus 4.8 runs roughly $5 and $25 for the same volume — a five- to six-times premium for whatever quality edge the frontier label still carries. At Coinbase's usage volumes, that difference isn't a rounding error. It's the entire savings.

The bigger lever, though, wasn't model choice at all. Caching hit rates went from 5% to 60% — the highest-leverage single change in the whole cost overhaul, ahead of routing to cheaper models. That detail is the real story: most of the money enterprises are burning on AI isn't going to smarter models. It's going to redundant computation that better plumbing would have avoided for free.

None of this reads as a retreat from AI. Token usage at Coinbase is growing exponentially, and 91% of engineers never hit the company's old usage ceilings — the caps simply stopped mattering once the per-token cost collapsed. Spend fell while consumption rose, the opposite of what "AI is getting expensive" coverage usually implies.

The part getting less attention than it deserves: Coinbase is a regulated, publicly traded financial exchange, and it is now routing an unspecified share of its workloads through open-weight models developed in China, decided by an automated system with no human in the loop and, so far, no public disclosure of what data touches which model. That's not a reason to suspect a Chinese model provider is doing anything untoward. It's a reason every fund backing fintech infrastructure should be asking portfolio companies which question is actually being answered when they tout model flexibility: cost, or governance.

Snowflake and Lindy have made similar moves toward open-weight Chinese models, and the direction of travel is clear: the frontier lab premium is becoming a discretionary purchase, not a technical requirement. That's good news for anyone building the orchestration, caching, and routing layer sitting above the models, because that's where the defensibility just relocated. It's a harder question for whoever is still pricing a product on which foundation model it uses as the pitch.