OpenAI has made its Decisions API generally available. The API takes a question and a set of options and returns an answer as a probability or a choice. According to a TechCrunch article dated October 22, the API became available to all customers that day. TypeSafe's "Jev" got to this market first, and OpenAI entered it not by building a new model dedicated to decisions but by reconfiguring its existing GPT-6 Luna. On standard input pricing, OpenAI charges $0.10 per million tokens against Jev's $0.042, so OpenAI is about 2.4 times more expensive. OpenAI trails Jev on price, and whether it can compete will likely depend on multimodal capabilities such as image and audio handling.
A decision-focused API built on GPT-6 Luna, now handling up to 400 questions per request
The Decisions API is designed to let AI make judgments on problems where the range of possible answers is set in advance. Developers define a question and its answer options, then supply text, images, or JSON as the basis for the decision. There are three response formats. "Predicate" returns the probability, from 0 to 1, that a specified condition is true. "Choice" selects one of the candidates and attaches a probability and a confidence level. "Score" calculates a probability-weighted average over a graded scale. OpenAI says the API processes requests up to 10 times faster than using GPT-6 Luna through the Responses API. That performance gain, however, is a figure supplied by the company itself.
The general release expands on the public beta. It adds integration with GPT-Live, prompt caching, and the ability to specify the inference region in the US or EU. The number of questions that can be processed in a single request has also doubled to a maximum of 400. Nikunj Handa, product lead for OpenAI's API division, told TechCrunch that the company did not build a new model tuned for decisions. Instead, it reconfigured Luna to speed up inference and cut costs. As a result, he said, response times fell from 1 to 3 seconds to 100 to 180 milliseconds. Handa said that, for now, Luna itself makes the decisions, though he also raised the possibility of dedicated training in the future.
Early adopters include Box, Datadog, Ramp, and Speak. According to Andrew Hsu, CTO of language-learning company Speak, processing response time dropped from 1.4 seconds to 0.13 seconds.
Input pricing: $0.10 for OpenAI, $0.042 for Jev
The key figure when comparing decision API pricing is the per-token input price. Neither OpenAI nor TypeSafe charges for output tokens, so standard usage costs depend on the input token price. OpenAI's official price list shows $0.10 per million tokens, while TypeSafe lists Jev at $42 per billion tokens. Converted to a per-million-token basis, that is $0.042, which means OpenAI's standard input price is about 2.4 times Jev's. The exact ratio is roughly 2.38, from dividing 0.10 by 0.042.
At an assumed rate of 150 yen to the dollar, that works out to about 15 yen per million tokens for OpenAI and about 6.3 yen for Jev. On its own website, TypeSafe claims that Jev's input price is 1/238 that of Claude Fable 5.1. OpenAI's $0.10 is low in absolute terms, but the same decision API market includes a competing service available for less than half that price.
The roughly 2.4x gap, however, applies only to list-price comparisons without caching. OpenAI supports prompt caching: when 1,024 or more tokens at the start of a shared input are cached across multiple questions, the cached input price drops to $0.01 per million tokens. That is below Jev's standard price, though writing to the cache incurs a charge. Jev's published price list, by contrast, shows only input and output prices, with no dedicated cache rate. For use cases that repeatedly rely on long shared instructions, actual costs could therefore flip in OpenAI's favor.
Performance rankings don't currently favor OpenAI either. According to TechCrunch, Jev ranks third on Artificial Analysis's "Decision Index," while OpenAI's model sits at 14th. TypeSafe's product is also reportedly the most used in OpenRouter's decision model category. Judging by published standard pricing and index rankings, OpenAI is the one chasing an established competitor.
For developers in Japan, the region where inference runs is also a consideration. The region-selection feature OpenAI announced covers the US and EU, while Asia only gets added capacity to reduce response latency. For companies that require data to be processed domestically, this announcement alone is not enough to conclude that their requirements are met.
Can images and audio set OpenAI apart? What to watch next
TechCrunch points to multimodal support as an area where OpenAI could build a competitive edge. On image classification, Handa said Luna performs strongly and also delivers good results on calibration, which measures how closely the model's stated confidence matches its actual accuracy. For audio, the general release adds GPT-Live integration. As an example use case, OpenAI cites a voice AI agent handling customer support that must decide mid-conversation whether to ask a clarifying question, route the inquiry, or hand off to a human agent.
It is worth noting, however, that Jev also supports image input. According to TypeSafe's blog, "jev-1.13.0" has had multimodal capabilities from the start, and image input can be used by specifying a dedicated header. This feature is offered as an experimental "danger zone" capability, and the company does not guarantee its use in production. TypeSafe also says that accuracy drops for every model when images are input rather than text. There is therefore no basis for concluding that Jev's image processing is inferior to OpenAI's. The evidence for OpenAI's advantage also rests mainly on the company's own explanations for now.
Response speed is also hard to compare directly. OpenAI cites 100 to 180 milliseconds, while TypeSafe cites 0.114 seconds per workflow, but each figure was measured under the company's own conditions. Without consistent measurement targets and processing conditions, these numbers cannot settle which service is faster.
Still, the reason OpenAI's multimodal support may offer room to differentiate is that the two companies deliver these features differently. OpenAI includes image input and GPT-Live integration in its general release, while Jev's image input remains an experimental feature without a production guarantee. That does not mean OpenAI has established a performance lead, but there is a gap in how far each offering is ready for real-world operation.
The conditions that would change this view are clear. If Jev's image input exits its experimental phase and becomes an official production feature, at least the difference in how image support is offered would disappear. And if OpenAI climbs from 14th to match Jev's 3rd place on the Artificial Analysis Decision Index, it would no longer be the one playing catch-up on that index's performance measure. Handa said that if response times fall to 60 milliseconds, it would become feasible to call the decision API on every frame of a video. Going forward, how much faster decision processing involving images can get, and what it actually costs in real-world operation, will be key indicators for judging the two services' competitiveness.