OpenAI's Decisions API Costs 2.4x Jev's Rate for AI Lead Routing

OpenAI opened its Decisions API to every developer as a public beta on October 6. Instead of writing text, it returns typed answers about text or images. An answer can be a probability that a statement is true, one choice from options you supply, or a score on a scale you define. Per OpenAI's developer guide, it runs only on GPT-6 Luna answers about 10 times faster than the same model through the Responses API. It costs $0.10 per million input tokens, with nothing charged for output. OpenAI expects general availability in the coming weeks.

Much of what a revenue team asks AI to do has this shape, like routing a lead or flagging the reply that asks for a meeting. A routing call that sends 1,000 tokens of lead record and rules costs a hundredth of a cent at OpenAI's rate, so 100,000 inbound leads a month come to about $10. That's our arithmetic, and it moves with how much of the record you send.

OpenAI entered the decision market as the expensive option

OpenAI didn't set that price. TypeSafe's Jev opened the category in mid-September and lists $42 per billion input tokens, which is $0.042 per million. Perplexity launched its own decision model on October 1 at $0.04, and its documentation now lists $0.02. By our math, OpenAI charges about 2.4 times Jev's rate and five times Perplexity's.

The $0.10 also matches what OpenAI already charges for Luna input through its regular API. The new endpoint drops Luna's $0.50 output charge, and it drops caching with it. A team that caches a long, fixed rubric through the Responses API could pay less there, as developers on OpenAI's forum noted within a day.

The price of a decision fell in the 16 days after Jev's launch, when Amazon, Cloudflare and Perplexity shipped cheaper or free versions. OpenAI sells the decision on contract terms a security team may already have approved. Its guide lists zero data retention and HIPAA support for eligible customers, plus data residency in the US and Europe. In the Hacker News thread on the launch, which drew 360 points, one developer said their team had just moved from OpenAI to Anthropic on AWS for a data-retention guarantee. For a company that already buys from OpenAI, the premium buys a decision endpoint that skips a new vendor review.

The model was never the costly part of AI routing

Early tests suggest the cheaper models are also the better ones. On OpenAI's forum, one developer reported that across a few hundred test questions, Jev was faster, cost a third to a half as much, and made fewer confident wrong answers on judgment calls. He called it an early signal, not a benchmark. Anthus, a consultancy that tunes Jev for clients, gave plain Luna 3,600 reasoning problems. When Luna said it was at least 99% sure, it was right 68% of the time. Anthus tested Luna through the regular API, not the Decisions endpoint, and it has a stake in Jev.

Calibration decides whether automated routing works, because the design is to act automatically above a confidence threshold and send the rest to a person. OpenAI's guide tells developers to set those thresholds from labeled examples in their own application. Someone on the RevOps team has to build that labeled set.

Jason Lemkin of SaaStr made the same point about Jev on LinkedIn. Decision models take “real work: testing, rewriting prompts, fixing processes,” he wrote. They also need clean data, because they can’t fill gaps in a thin record the way Claude can. Lorcan O'Rourke, a solutions engineer at Clay, posted a demo of 4,000 Jev decisions across 1,000 accounts in 40 seconds for five cents. It worked because Salesforce history, warehouse usage and call data were already sitting on each account record in Clay. At OpenAI's rate it would cost about 12 cents. The costly part was getting the data there. As our piece on Salesforce's agent meter argued, that margin goes to whoever owns the record.

What this means for RevOps leaders

Start with what you already pay. Salesforce has listed an Agentforce action at $0.10 since May 2025, about a thousand times the OpenAI token cost of the routing call above. An action does more than decide, since it can update a record or trigger a workflow, so the two aren't like for like. The gap still gives you a question for every AI scoring or routing line item in your stack: what does the decision step cost the vendor, and what is the rest of the price paying for? A good answer points to data access and accuracy on your hardest cases. A weak one points to the model.

Then test before you build. Run a few hundred routed leads with known outcomes through Jev, Perplexity and OpenAI. Check accuracy on the cases your team argues about, and check whether 90% confidence is right nine times in ten. The token bill for that test rounds to zero. Labeling the cases is where the cost sits.

My bet is that OpenAI adds caching or cuts the Decisions API price before general availability, because its own forum is already pricing it against Jev. If it holds at $0.10, its wins will come from companies where the security team chose the vendor.

Calibration decides whether automated routing works, because the design is to act automatically above a confidence threshold and send the rest to a person.

Run a few hundred routed leads with known outcomes through Jev, Perplexity, and OpenAI, check accuracy on the cases your team argues about, and check whether 90% confidence is right nine times in ten.

By Business of GTM