Decision models accounted for 26% of Vercel AI Gateway requests but just 0.53% of spending

Your service receives a customer inquiry and has to decide whether to send it to support, pass it to sales or ask for clarification. That does not always require a large language model to handle an entire conversation. Sometimes, a brief decision from a handful of options is enough. According to Vercel's November report, these tasks already account for 26% of AI Gateway requests while representing just 0.53% of estimated spending.

That is a useful reason to revisit the architecture of an AI product. If a model writes replies, sorts inquiries and chooses follow-up actions, it is worth measuring the cost of each stage separately. Some of that work may be suitable for cheaper, specialized models.

Brief decisions have become a distinct traffic category

Vercel uses the term decision models for models that classify data, assign scores or choose between options. Instead of a detailed response, the application receives a brief result it can use in the program's next step. The company estimates that the average request to these models costs roughly 70 times less than a language model request.

Consider an online store. A customer writes: "My order hasn't arrived. I want a refund." First, the system needs to identify the type of inquiry. It can then send the message to the appropriate handler, retrieve the order details and only then prepare a substantive reply. Text generation stays where it adds value, while routing becomes a separate operation.

In October, nearly a third of the gateway's paying teams used these models. Among teams processing more than 10 billion tokens per month, adoption reached 70%. At high volumes, even a small saving on each action can make a substantial difference to the total bill.

But there is an important detail: almost all traffic in this category went to Jev from TypeSafe AI. The 26% figure therefore primarily reflects heavy use of one particular product within one particular gateway. It does not yet prove that dozens of independent providers have found mass-market demand.

Anthropic attracts new teams, while OpenAI captures more spending

OpenAI and Anthropic were nearly tied in their reach among paying teams: in October, 55% and 54% of teams used their models, respectively. Those shares overlap, since a team can use both companies.

Among new teams, Anthropic became the primary provider almost twice as often as OpenAI. The definition matters here: Vercel considers a team's primary provider to be the one that received the most language model requests from that team in its first month. This does not mean an exclusive contract or a final choice of technology stack.

Meanwhile, OpenAI's share of gateway spending rose from roughly 9% in December to 15% in October. Anthropic's share roughly halved over the same period, to 32%, although it retained the lead on this metric. Among the top 1% of teams by spending, more than 90% use both providers.

For founders, the practical takeaway is that popularity at initial adoption and the volume of paid work are different metrics. One cannot reliably be inferred from the other.

Models change faster than product architecture

More than half of October's language model requests went to versions that had been available on the gateway for less than 60 days. Around half of teams processing at least 10 million tokens changed their primary model during the month. Of those that switched, 60% moved to a different lab.

That level of movement makes it useful to maintain your own set of evaluation tasks and the ability to switch providers without reworking business logic. Before replacing a model, compare quality on real inquiries, latency and the cost of a successful outcome. A cheap request does little good if it requires manual review afterward.

The report has limitations. It describes anonymized Vercel AI Gateway traffic for October, with spending estimated using public pricing. Team size is defined by token volume; the text provided does not state the absolute number of teams. There is also no data on decision quality or comparable latency.

A sensible next step, then, is to choose one frequent classification task, collect real examples and compare a specialized model with the current solution. That test will show whether the gateway's impressive statistics translate into savings for your product.