OpenAI’s decision API costs more than Jev, but adds image support

OpenAI lists its Decisions API at $0.10 per million input tokens, compared with $0.042 for TypeSafe AI’s Jev. Neither charges for output tokens. The competing services target routine judgments inside software, where the cost of each call matters—but speed, supported inputs and accuracy also determine whether a cheaper decision is useful.

These are jobs such as sorting documents or routing customer messages, rather than writing a reply. At thousands of judgments a day, small differences in delay, expense and error rates can change whether automation is worthwhile.

Jev made the narrower job explicit

TypeSafe introduced Jev on September 15, 2026, in early access. Its launch post described a model built to return structured decisions that software can use directly, giving up free-form text generation. The possible outputs and their structure are defined before the model answers.

A September 29 evaluation paper describes those answers as choices from fixed options, positions on a scoring scale, or probabilities that a statement is true. That distinction puts the decision—not a written explanation—at the center of the interface.

OpenAI tuned the interface, not the model

Fortune reports that OpenAI rolled out Decisions API on October 6. OpenAI’s Nikunj Handa credited Jev with inspiring the service in a Latent Space interview. For its first version, the team used GPT-6 Luna without further training, constrained its answers, processed multiple questions together and tuned delivery of the first decision.

OpenAI says the dedicated interface can return decisions up to ten times faster than calling Luna through its general-purpose Responses API. That is a comparison between two ways of accessing the same model, not a claim that Decisions API is ten times faster than Jev.

$0.10OpenAI Decisions API

OpenAI’s listed input rate; output tokens are free.

$0.042TypeSafe Jev

Jev’s listed input rate; output tokens are free.

Both services can choose among preset options, estimate whether a condition is true and assign scores. OpenAI also accepts images, allowing an application to evaluate a photograph or browser screenshot. Jev currently accepts only text, so the input format can determine which service fits a task.

The answer still has to be right

Developers still need to establish whether the judgments are accurate enough for their particular task. In an email-sorting experiment, developer Aaron Roy asked Jev to flag messages needing his attention. Important emails slipped through. Specifying money, children and medical issues reduced the misses in his test.

Confidence thresholds also shape what gets automated. A customer-service application could route messages above a chosen confidence level and send uncertain cases to a person. The practical next step is testing those boundaries on the actual work: a fast, inexpensive classification is useful only if the workflow handles its mistakes.