In October, decision models took one in four requests passing through Vercel's AI Gateway, and they accounted for 0.53 percent of that traffic's cost. The same Vercel data keeps the two numbers apart on purpose, because those calls were fast and they were cheap, about a seventieth as much per request as a language model that writes text. That split is the concrete fact every operator should start from.
The reason those numbers matter is that the tooling has now arrived in force. Google launched Gemini Decision 1 with a dedicated Decisions API on November 11, following OpenAI, which opened its Decisions API to every customer on October 22, and Microsoft, which built its model on Alibaba's Qwen. All three charge $0.10 per million input tokens or less, and none charge for output. A decision model does not write prose. You give it state and a fixed list of choices, and it returns the best choice with a probability, in milliseconds. TypeSafe AI made the category with Jev, released September 15, and Vercel called it the fastest-adopted model it has ever measured.
What the adoption curve is telling you
Nearly a third of Vercel's paid teams used a decision model in October. Just over one in five teams running fewer than 10 million tokens a month used one, and seven in ten of the teams running more than 10 billion did. The teams under the most pressure to hold latency and cost flat adopted first, and about 11 percent of paid teams went further still, using a decision model while sending no text-generating request at all.
The volume is what lands. Almost all of it went to Jev, which alone took nearly four times as many requests as all of Anthropic's models combined. For every $1 a team spent on Jev, it spent about $64 on Anthropic. The small automated decisions cluster around the cheap model, while the expensive model still earns its keep on the work that does not scale that way.
The test, applied
The title's question asks how many of the small decisions inside your stack can honestly go to a model that answers faster than a chatbot. The answer is not a percentage you can look up, but a test you can apply to each one: hand the call over when the decision is cheap to get wrong, well defined, and high in volume, and keep it when it is expensive to get wrong.
The cheap and easy decisions are everywhere in a typical stack. Classifying an incoming request into a queue, routing it to the right team, deciding whether a prompt is safe, and gating whether the next step should run at all. Each is a choice from a fixed list, exactly what a decision model is built to return, and each would otherwise slow a chatbot and cost you for every word it writes.
The payoff shows up in numbers operators are already seeing. OpenAI's product lead Nikunj Handa said the decisions API now answers in 100 to 180 milliseconds, down from one to three seconds, and a company called Speak reported cutting the time between actions from 1.4 seconds to 0.13 seconds. That speed only matters if the decision would otherwise sit behind a chatbot, and most of the small ones do.
The signal that a decision is a good candidate lives in both the data and the judgment. Vercel's own product team is building the judgment into the gateway. In October they added confidence-based fallbacks, letting a decision request escalate to a different model when the primary answer falls below a confidence threshold you set. The model is allowed to shrug, so an unsure answer routes up to a stronger model instead of being forced into a cheap commitment.
Keep the hardest calls where they belong
The instinct to hand everything over has a limit, and the benchmarks make it honest. On Artificial Analysis' Decision Index v2, the best decision models score about three points below the language models. Claude Haiku 5.5 tops the index at 77, GPT-6 Luna sits at 76, both ahead of the leading decision model, Drex Max, at 73. The price gap is steeper than the score gap, with GPT-6 Luna at $1,530 per million tasks versus $102 for Drex Max, and 11.3 seconds per task against 275 milliseconds.
So the language model buys accuracy and handles ambiguity, at 15 times the cost. Whether that is worth it depends entirely on the decision. For a content-moderation check that fires on most of your traffic, the cheap model is almost always the right call. For a judgment call that a customer will appeal, the extra points may be worth the extra money. The operator's job is to know which side of the line each decision falls on, and to route by it.
Trust the number, but read it right
TypeSafe's own launch forced developers to confront the probability that comes with a decision model. Armin Ronacher of Earendil put it to TechCrunch as a question of where to draw the line, noting that a 50 percent answer is a coin toss you would be wise to ignore, while a 95 percent answer is one you can act on. The model hands you a number, and deciding how much to trust it is an operator's decision, not a detail to defer to an engineer later.
There is also the honest limit of the benchmarks. A recent benchmark of 28 decision models, MMDM, showed the best model scoring only 36.81 percent on hard questions with several right answers. Models that look sharp on easy cases can struggle when the right answer is less obvious or the options are close. Forty-five percent of the index now uses private data, up from a quarter, precisely to resist benchmark gaming, which is itself a sign that the public numbers can overstate real-world performance.
The answer to the title
You can honestly hand the call to a decision model for the high-volume, well-defined, cheap-to-get-wrong decisions, and you should. The Vercel data shows this is already the norm for teams running AI at scale, and the economics are hard to argue against. Keep the language model for the decisions that are expensive to get wrong, and build a confidence gate between the two so an uncertain answer routes up instead of being forced into a cheap one.
The three labs now each sell you a model that answers faster than a chatbot, and they are fighting over the small calls. Your job is simply to decide, decision by decision, which of those small calls can afford to go cheap.
Sources
- Google is the latest tech giant to chase TypeSafe's 8-week-old Jev | AI Weekly #543 (AI News Weekly, 2026-11-13)
- Decision models take 26% of requests; Anthropic now wins new teams 2:1 (Vercel, 2026-11-09)
- AI Gateway adds confidence-based decision fallbacks (Vercel, 2026-10-06)