A narrower API for a specific job
OpenAI has released the Decisions API in public beta, a dedicated endpoint built specifically to return typed answers to developer-defined questions over text and image inputs, rather than the open-ended prose a standard chat completion produces. Three question types are supported: predicates that return a probability a given condition holds, such as detecting damage in a product photo; choices that pick one option from a fixed set with per-option confidence scores, such as routing a support ticket to billing or technical teams; and scores that return a probability-weighted rating across ordered severity levels.
The endpoint runs exclusively on GPT-6 Luna, the reasoning model OpenAI released alongside GPT-6 Sol in September, and OpenAI says it is up to ten times faster than calling the same model through its general-purpose Responses API. That speed difference is the actual product here: for workloads that only need a routing decision or a classification, paying for and waiting on full generative output is pure overhead the Decisions API is designed to strip out.
A pricing model built for volume
Pricing is $0.10 per million input tokens, with no charge for output tokens, cache reads, or cache writes, which is a meaningfully different economic model than a typical generative API where output tokens usually cost more than input. For a workload like content moderation, support ticket routing, or product photo inspection running millions of times a day, that structure changes the cost calculus significantly compared to running the same classification through a general chat endpoint and parsing a text response afterward.
The target use cases OpenAI lists, content classification, request routing, deciding an agent's next action, product photo inspection, and relevance or severity scoring, all describe exactly the kind of high-volume, low-latency decision points that sit inside larger automated workflows rather than user-facing conversations. This is infrastructure aimed at the plumbing of enterprise automation, not at a new chatbot experience.
The competitor this move walked over
Three weeks before OpenAI's release, a startup called TypeSafe launched a product called Jev, described as a System One decision model, backed by a $40 million seed round led by DCVC. Jev staked out the same niche, fast, cheap, typed decisions, at a price of $0.042 per million input tokens, under half of what OpenAI now charges for the same category of task through the Decisions API.
Informal developer testing posted to OpenAI's community forum, based on a few hundred test questions, found Jev faster at roughly 0.3 seconds versus 1.6 seconds median response time, cheaper, and more accurate on nuanced judgment calls, with the two products running roughly even on simple factual checks. Jev's clearest gap was that it did not yet support image input, the exact feature OpenAI built into the Decisions API from day one, prompting one developer to post bluntly on social media: image support means Jev is finished.
Why this is a build-versus-buy case study, not just a product launch
This sequence, a well-funded startup launches a sharply focused product, and a platform incumbent ships a broader version of the same idea within weeks, is one of the clearest live examples enterprise buyers will get this year of the risk in betting infrastructure strategy on a narrow AI startup rather than a platform player's adjacent capability. Jev's technical advantages on speed, price, and accuracy for simple decisions are real and documented, but they may not matter if OpenAI's distribution and existing customer relationships make the slightly slower, pricier option the path of least resistance for most buyers.
It is also worth noting that OpenAI was not alone in moving into this space. Reporting on the launch names Liquid AI, AWS, Perplexity, and Cloudflare as having released comparable decision models or APIs within roughly the same week in late September, which suggests specialized, low-latency decision inference is becoming a category multiple large platforms are racing to own simultaneously, not a niche OpenAI spotted first.
What enterprise technology leaders should actually do with this
For any enterprise currently routing classification or content-moderation workloads through a general-purpose chat completion and parsing text output, the direct recommendation is to benchmark the Decisions API, and comparable narrow products like Jev, against your actual production traffic before committing to either, since the real-world tradeoffs on latency, cost, and accuracy will vary by workload in ways a few hundred informal test questions cannot fully predict.
The bigger strategic lesson is about vendor selection discipline: when evaluating any infrastructure startup whose pitch is a narrow, well-executed slice of a larger platform's eventual roadmap, assume the platform player will eventually ship a comparable capability, and price that risk into the contract terms, not just the product evaluation. Jev's team built something technically credible in weeks and still got boxed in by a feature gap within a month of launch, which is the exact scenario every CTO evaluating a point-solution AI vendor should be modeling before signing a multi-year commitment.
