Decision APIs: Fast, Cheap AI Judgements and Where They Fall Short

A field guide to the decision APIs released in September and October 2026, written for leaders who have to decide where AI should make the small, repeated calls inside their workflows and agents, and where it should not.

In one paragraph

A decision API answers a question you have already written the answers to. "Is this complaint urgent?" "Which team handles it?" "Should the agent call this tool?" You send the context, the API returns a typed answer with a probability, usually in a few hundred milliseconds and for a fraction of what a language model costs. TypeSafe AI started the category with Jev in September 2026. OpenAI followed with its Decisions API, in public beta since 6 October, and Liquid AI, Cloudflare, Perplexity, AWS and Microsoft all shipped their own within weeks. For a business, they are a cheap way to put judgement into the if-statements of a workflow. They are not yet a way to make decisions you would have to explain to a customer or a regulator.

What a decision API is, and what it is not

TypeSafe calls these "System One" models, after Kahneman's fast, intuitive mode of thinking, and describes Jev as "unstructured state in, typed probabilistic decisions out." The trade is simple: the model gives up writing text, and in return it is fast, cheap and always answers in the shape you asked for.

That makes it different from a language model. OpenAI's own documentation points you to Structured Outputs or function calling when you need custom JSON, an explanation or a tool call. A decision API only picks, scores or estimates.

It is also less new than it looks. An AWS builder article traces the lineage through BERT classifiers, zero-shot inference models, reward models, LLM-as-judge, Llama Guard and model routers, and concludes that Jev changed "the packaging … more than the parts." The packaging matters: a general-purpose judgement behind one API call, with no training, is what earlier classifiers could not offer.

One note on the name: this has nothing to do with business-rules engines or the DMN decision-table standard. Those execute rules you write. Decision APIs make a judgement from context.

How it works, for a non-engineer

You send three things: the model, the input (text, or text and images), and a list of questions. OpenAI's endpoint is POST /v1/decisions and supports three question types:

• Predicate: the probability that a statement is true. "This email is a complaint": 0.91.
• Choice: one option from a list you supply, with a probability for each. "Route to: billing, technical, sales, other."
• Score: a position on ordered levels you define, such as low, medium, high, critical. The score is a probability-weighted average, so it can fall between levels.

One request can carry many independent questions. OpenAI's Python SDK types allow 1 to 200 questions, 2 to 255 options per choice and up to 128 images per request. Questions that depend on each other's answers need separate requests. gpt-6-luna is the only model, and the guide says images must be sent inline rather than by URL.

TypeSafe's equivalent uses a /v1/systemone endpoint with choice, score and noul (yes/no) questions. That format has become an informal standard: Ollama implemented it, Microsoft Foundry uses it and Cloudflare describes its models as "fully Jev-API compatible". OpenAI's format differs, and one independent tester found that "switching providers needs a small adapter, not just a new endpoint URL."

Who has released what

Developers moved quickly. Vercel reported that nearly 13% of paid teams on its AI Gateway were using Jev within 24 hours, making it "the fastest-adopted model in AI Gateway history."

What they are good for

OpenAI's guide names three jobs: classify content, route requests and prioritise work. Its examples are checking product photos for visible damage, routing complaints to the right department, rating the severity of an issue and, since a 15 October changelog entry, choosing actions from live voice.

In practice the strongest uses fall into four groups:

• Triage and routing: inbound email, tickets, claims and documents sorted to the right queue with an urgency score, and the uncertain ones sent to a person.
• Agent guardrails: a fast check before an agent acts. Is this tool call within policy? Does this reply leak personal data? TechCrunch reported one start-up checking every agent action with Jev for about $2.94 where a frontier model would have cost about $372, figures TechCrunch did not independently verify.
• Agent next steps: choosing which tool, model or sub-agent should handle a request, in the time a user will not notice.
• Visual inspection: pass or fail on product, damage or quality photos. Liquid AI reports 85 to 97% on an industrial anomaly dataset, from its own tests.

Screening people is also on the list. TypeSafe cites a candidate-screening customer, Jack & Jill, reporting an 88% cost reduction. That is a vendor claim, and it is also the use most likely to fall under employment and data protection law, covered below.

We use decision models in production ourselves. At JobVantage, the recruitment intelligence platform we built and run, they are part of the matching pipeline on Hexploits Cloud, helping keep matches accurate and highly relevant while keeping the time to return them as low as possible.

The benefits

• Speed. OpenAI says its endpoint is "about 10x faster than the Responses API", without publishing the method. Independent testers measured mean latency of 144 to 194 milliseconds on a banking classification task.
• Cost. You pay only for input. OpenAI charges no output or cache fees, and the rivals are cheaper still.
• Answers that always fit. The output is always one of your options, so there is no parsing, no retry for malformed JSON and no invented category.
• A number to set policy on. A probability lets you choose a threshold: act automatically above 0.9, send to a person between 0.6 and 0.9, reject below. That is a design decision a manager can own.
• No training data to start. You write the options and their descriptions. A trained classifier needs thousands of labelled examples first.

What does it cost at volume? As our own arithmetic from list prices: a million decisions at 500 input tokens each is 500 million tokens. That is about $50 on OpenAI ($55 with EU data residency), $21 on Jev, $20 on Liquid d1 and $5 on Perplexity's flash model. Images, long inputs and regional premiums add to that, but the order of magnitude is pounds per million decisions, not thousands. At steady high volume there is a second option: an open-weight decision model on a dedicated GPU, where the cost is a fixed monthly figure however many decisions you make. Which is cheaper depends on your volume, so model both.

The drawbacks

• Overconfidence. The Artificial Analysis Decision Index found that on knowledge and reasoning tasks "accuracy falls below that probability for all decision models." When a model says 0.95, it is right less often than that.
• Sensitivity to how you ask. Developers on the Hacker News launch thread reported that changing the order of options changes the probabilities. Perplexity's documentation warns that grouping questions in one request changes the answers.
• Refusals without reasons. One tester found OpenAI's endpoint returning a refusal with "no reasons or probabilities" on ordinary banking queries. Your code has to handle a missing answer.
• No explanation. A probability is not a reason. If a customer asks why, the API cannot tell you.
• Moving targets. OpenAI's endpoint is in beta, with general availability expected "in the coming weeks", and it has published no policy on pinning a model version. TypeSafe writes of its own pricing: "We can't prove it isn't subsidized."
• Prompt injection. A decision model reads untrusted text like any other model. A cleverly worded email can still push it towards the answer an attacker wants.

What they can and cannot do

They can:

• answer yes or no, choose one of up to 255 options, or score on a scale you define
• read text and, for OpenAI, Liquid and Cloudflare, images
• answer many independent questions about the same input in one call
• return a probability you can set thresholds on
• run on your own hardware or dedicated EU GPUs, if you choose an open-weight model such as Clef, d1-3B or Strands Decider

They cannot:

• write text, summarise or explain their answer
• answer questions that depend on each other in a single call
• take files, audio or tool calls (OpenAI's endpoint)
• process data inside the UK (OpenAI's endpoint)
• guarantee that a stated probability matches how often they are right
• reliably beat a classifier trained on your own labelled data for a stable task

Benchmarks, and how to read them

Most published numbers come from the vendors. Three sources are worth knowing.

The Artificial Analysis Decision Index, published on 13 October 2026, is the first independent leaderboard. It weights four tests equally: knowledge and hallucination, multiple-choice reasoning, deciding when to call a tool, and safety classification.

Two cautions. The GPT-6 Luna row is the language model, not OpenAI's Decisions endpoint, which was not separately listed. And Liquid AI says its d1-ultra, released two days later, leads the index, a claim Artificial Analysis had not published at the time of writing.

A banking-intent test by StackToHeap (8 October) ran 770 questions from the public BANKING77 dataset through both OpenAI and Jev:

OpenAI was faster and better calibrated, Jev slightly more accurate and cheaper per run. The small trained classifier beat both. In the author's words: "Neither hosted API displaced it on this fixed taxonomy."

A logic-puzzle test on GitHub (10 October) scored Jev at 81.7% and OpenAI at 61.2%, with a median OpenAI latency of 221 milliseconds and 17 questions OpenAI refused to answer. It uses puzzles from Jev's own benchmark, which may favour Jev.

The lesson is not which vendor wins. It is that results swing by task, and the only benchmark that matters is a few hundred of your own cases with known answers.

Data protection and regulation

Where your data is processed. OpenAI's data controls page lists regional processing for the US and for Europe (EEA and Switzerland). The UK endpoint offers storage in the UK but no processing there. Using any non-US region requires approval for OpenAI's abuse-monitoring controls, and data residency carries a 10% price uplift. Zero data retention is available to eligible customers, with exceptions for prompt caching and images. If processing must stay in the UK, or on infrastructure you control, the open-weight models are the route.

That is what we built Hexploits Cloud for. It runs open-weight models on physical NVIDIA RTX PRO 6000 Blackwell GPUs reserved to a single customer, in our eu-central datacentre in Germany. Models stay resident in GPU memory, prompts and data stay in the EU by default, there is no per-token metering or rate limit, and the cost is a fixed monthly figure. For a decision model, that means the data never reaches the model vendor and the version only changes when you change it.

Automated decisions about people. Since 5 February 2026, Articles 22A to 22D of UK GDPR govern significant decisions based solely on automated processing. They allow such decisions with safeguards: telling people, letting them make representations, and offering human intervention and a way to contest. A decision API that rejects a loan application, a claim or a candidate without a person in the loop is exactly that kind of decision, whichever vendor supplies it.

The EU AI Act. If you serve EU customers, using any AI system, decision model included, to score creditworthiness, price life or health insurance or screen job applicants is high-risk under Annex III. We set out the obligations for finance and recruitment in earlier insights.

Explaining outcomes. The FCA's Consumer Duty expects firms to evidence good customer outcomes. A probability with no reason attached makes that harder, so log the input, the question, the options, the probabilities and the action taken for every decision that affects a customer.

No regulator has commented on decision APIs specifically. The rules above apply because of what the decision does, not how it is made.

Decision API, language model or your own classifier?

Many systems will use all three: a decision model to route and gate, a language model where words or reasons are needed, and a trained model where the task is stable and the volume is high.

Timeline

What to do now

1. List the if-statements. Find the places in your workflows and agents where a person or a language model makes the same small judgement many times a day.
2. Build an evaluation set before choosing a vendor. Two to five hundred real cases with known answers, tested for calibration as well as accuracy.
3. Set thresholds and a human-review band, and log every probability alongside the action taken.
4. Decide where data may be processed. For UK personal data, that may mean the EU endpoint with approval, or an open-weight model on your own infrastructure.
5. Put an adapter in front of the vendor, so that a better or cheaper model is a configuration change, not a rewrite.
6. Treat decisions about people as automated decisions, with the notices, review and contest routes UK GDPR requires.

How Hexploits helps

We build decision support that scores, ranks and routes against your own history, with thresholds a manager can understand and the evaluation set to prove them. We put fast checks and approvals into agentic workflows, as we did in swarmd.ai, and we build document and email processing that classifies and routes with a person in the loop. Our own product, PII Purge, uses LLMs to classify and redact personal data in production on EU infrastructure, and we build MCP servers that put products like it inside AI assistants. We run decision models in production for JobVantage's candidate matching. Where data must stay under your control, we deploy open-weight models privately on Hexploits Cloud, on dedicated GPUs in the EU at a fixed monthly cost, and we test the result for prompt injection and misuse.

Request a proposal or talk to an engineer. You will have a written scope and an indicative price within two working days.