Machine Learning
OpenAI's Decisions API vs TypeSafe's Jev: "System One" Models and the Race for Fast, Typed AI Decisions
A Facebook post asked whether OpenAI just made a better Jev. We compare OpenAI's new Decisions API with TypeSafe's Jev, two models that return typed, probabilistic decisions instead of text: context size, images, price, compliance, and why calibration is the machine-learning problem underneath.
Akmal Alif · 10 October 2026 MYT
On 7 October 2026 Shee Tze Jin, who writes a daily AI trivia series on Facebook, posted a short comparison under a pointed headline: "OpenAI Just Made a Better Jev" (Shee Tze Jin, 2026). The day before, OpenAI had released its Decisions API in public beta. Three weeks earlier, a startup called TypeSafe AI had launched Jev, a model built to do almost the same job.
Both products answer one kind of question: not "write me something", but "decide this". Is this transaction fraud? Which team should handle this ticket? How severe is this issue? This post explains the idea behind them, checks the numbers in Jin's comparison, and looks at why a calibrated decision is a different machine-learning problem from generating text.
Jin's argument in brief
Same idea. You give the model context and a focused question, and it returns a structured decision.
OpenAI takes far more input. Jev handles about 64,000 tokens; OpenAI's model takes about 1.05 million, so you could include a long document, a whole conversation history or hundreds of pieces of evidence before asking for one decision. OpenAI also accepts images.
Jev is cheaper. $0.042 per million input tokens against OpenAI's $0.10, so OpenAI costs about 2.4 times as much per token.
But for simple decisions, the price gap barely matters. If you need enterprise features such as zero data retention and logging, or image input, OpenAI's compliance and reliability may matter more than the price of a new startup's model.
What a "System One" model is
The name comes from Daniel Kahneman's Thinking, Fast and Slow. System 1 is fast, automatic judgement; System 2 is slow, deliberate reasoning (Kahneman, 2011). TypeSafe uses the label for a class of models built for fast decisions inside software, not for conversation (Almeida, 2026).
Most software already makes thousands of small judgements: route this email, flag this payment, score this lead. Today a developer who wants an AI to make them usually prompts a chat model, asks for JSON and then parses and validates whatever text comes back. A System One model removes those steps:
Typed output. The answer is a probability, a choice from a fixed list or a score, never free text.
Probabilities included. Each answer carries a probability or confidence, so the program can set its own threshold.
Speed and cost low enough to call the model inside ordinary code paths, like a smart if-statement.
Jev: the startup's version
TypeSafe AI's founder, Diogo Almeida, announced Jev on 15 September 2026; the blog post now carries an updated date of 9 October (Almeida, 2026). Almeida writes that they previously worked at OpenAI on methods that made language models better at following instructions. Jev is named after the economist William Stanley Jevons. TypeSafe's argument is that each fall in the cost of intelligence, as with coal in Jevons's day, unlocks more uses than it saves (Jevons, 1865).
What TypeSafe says about Jev (Almeida, 2026; TypeSafe AI, n.d.):
A new architecture. A parallel sampler produces the whole answer in one step instead of token by token, and the model gives up free-form text entirely.
A new training objective. Reinforcement Learning for Calibrated Decisions (RLCD) rewards answers whose probabilities are well calibrated, instead of rewarding human preference or verifiable correctness alone.
Claimed speed and cost. 70–500 ms per decision, which TypeSafe puts at 40–200 times faster than frontier language models on this kind of query, at $0.042 per million input tokens with free output.
Limits. 64,000 tokens per request, text only, and up to 255 choices before a slower two-stage method kicks in.
TypeSafe lists its own caveats: speed was measured from laptops on the US West Coast, pricing may be subsidised, and its comparisons come from internal workflow tests rather than public benchmarks. Its claim that Jev "can't hallucinate" means its answers cannot break their type. A well-typed answer can still be wrong.
The Decisions API: OpenAI's version
OpenAI's changelog records the Decisions API's release in beta on 6 October 2026 (OpenAI, 2026). The guide describes it (OpenAI, n.d.-a):
One endpoint, one model. Requests go to a single endpoint and run on gpt-6-luna, which OpenAI says is about ten times faster than its general Responses API.
Three question types. A predicate returns the probability that a condition is true. A choice returns one option from a list, with probabilities and a confidence. A score rates input against a rubric of levels, returning a probability-weighted score that can fall between levels.
Evidence. Inputs can be text and images, with images sent inline. Several independent questions can share one request.
Price and compliance. $0.10 per million input tokens, with no charge for output or caching, plus zero data retention, HIPAA support and processing in the US or Europe for eligible customers.
The underlying model, GPT-6 Luna, has a 1.05-million-token context window, accepts text and images and has a knowledge cutoff of 18 May 2026 (OpenAI, n.d.-b).
Checking the comparison
Claim in Jin's post | What the documentation says | Verdict |
|---|---|---|
Jev handles about 64K tokens | 64K tokens per request, of which 32K for the input and the longest question | Correct |
OpenAI handles up to 1.05M tokens | GPT-6 Luna's context window is 1.05M, with up to 922K of input; the Decisions guide gives no separate limit | Correct for the model |
OpenAI accepts images | Text and image input; Jev is text only | Correct |
Up to 128 images per request | We found no such limit in the Decisions guide or the model page | Not confirmed |
$0.042 vs $0.10 per million input tokens, about 2.4 times | Both prices match, and both charge nothing for output | Correct |
OpenAI made a "better Jev" | Jev launched three weeks before the Decisions API | The order is right; "better" depends on the job |
Jin's post does not compare speed, and the published claims cannot be compared directly: TypeSafe measures against frontier chat models, and OpenAI measures against its own Responses API.
The machine-learning point: calibration
A typed answer is only half the product. The other half is the probability attached to it, and that number is only useful if it is calibrated: across all the times the model says 80%, it should be right about 80% of the time.
Modern neural networks are often overconfident (Guo et al., 2017). OpenAI's own GPT-4 report showed that its pre-trained model was well calibrated, and that post-training to make it a better assistant made calibration worse (OpenAI, 2023). That is the gap both products are aiming at. TypeSafe trains for calibration directly with RLCD. OpenAI's guide tells developers to set thresholds on their own labelled examples, weighing the cost of a false positive against a false negative.
That advice applies to both. Before trusting either model's 0.92, measure its calibration on your own data.
Jevons, and who wins
The Jevons argument is the most interesting part of TypeSafe's pitch. At a few cents per million tokens, a model call becomes cheap enough to run on every ticket, transaction or log line, not just the hard cases. Total spending on AI decisions may go up as the unit price goes down.
Who captures that spending is the open question. Jin's answer, compliance and reliability, explains why a startup can see a big platform ship its product category within weeks and still face a hard sale. TypeSafe's answer is a purpose-built model that is cheaper and, it says, faster. A practical way to keep that choice open is to wrap decision calls in your own small interface, so you can switch provider without rewriting the code that uses the answers.
References
Almeida, D. (2026, September 15). Introducing System One Models & Jev (Updated October 9, 2026). TypeSafe AI. https://typesafe.ai/blog/introducing-system-one-models-and-jev
Guo, C., Pleiss, G., Sun, Y., & Weinberger, K. Q. (2017). On calibration of modern neural networks. In Proceedings of the 34th International Conference on Machine Learning (Vol. 70, pp. 1321–1330). PMLR. https://proceedings.mlr.press/v70/guo17a.html
Jevons, W. S. (1865). The coal question. Macmillan.
Kahneman, D. (2011). Thinking, fast and slow. Farrar, Straus and Giroux.
OpenAI. (2023). GPT-4 technical report (arXiv:2303.08774). arXiv. https://doi.org/10.48550/arXiv.2303.08774
OpenAI. (2026). Changelog [Entry of October 6, 2026]. https://developers.openai.com/api/docs/changelog
OpenAI. (n.d.-a). Decisions [API guide]. Retrieved October 10, 2026, from https://developers.openai.com/api/docs/guides/decisions
OpenAI. (n.d.-b). GPT-6 Luna [Model page]. Retrieved October 10, 2026, from https://developers.openai.com/api/docs/models/gpt-6-luna
Shee Tze Jin. (2026, October 7). Jin Daily AI Trivia: OpenAI just made a better Jev [Facebook post]. Facebook. https://www.facebook.com/omega6jin/posts/pfbid0QTjWfqGzNtJpfjccZFFhmJXgGMmQgQE5QMzvtu8DzxxgS1pbhupFdyGzTaaLq8NVl
TypeSafe AI. (n.d.). Models. Retrieved October 10, 2026, from https://docs.typesafe.ai/models