Models that answer without writing

Between 15 September and 15 October 2026, a new kind of model went from one startup’s launch to products from OpenAI, Microsoft, Cloudflare, Liquid AI and Perplexity. Decision models do not generate text. You send an input and a set of typed questions, and the model returns a probability for each allowed answer. Most vendors bill input tokens only, and they quote latencies in tens to hundreds of milliseconds. An independent comparison by Artificial Analysis, published on 13 October, confirms the speed and cost advantage. It also finds that decision models trail language models on knowledge and reasoning, and that all of them were overconfident.

Scope: models and APIs that return probabilities over a fixed answer set instead of generated text. Period: 15 September to 17 October 2026. Performance figures are the vendor’s own unless stated otherwise. One later change, OpenAI’s general release on 22 October, is noted in its item.

15 September: TypeSafe launches Jev and names the category

• What happened: TypeSafe released Jev in early access and called it the first “System One” model, after the fast, intuitive mode in Daniel Kahneman’s Thinking, Fast and Slow. Software sends state plus typed questions. Jev returns typed answers with calibrated probabilities and supports up to 255 choices per question. TypeSafe reports end-to-end responses of 70 to 500 ms and prices input at $0.042 per million tokens, with output free (TypeSafe).
• Why it matters: the launch post set the pattern the rest of this list follows: a fixed answer set, a probability per answer, billing on input only. The post also lists its own caveats. Pricing may be subsidized, and its comparisons wrapped the language models in a layer that forced structured answers, which made them slower and more expensive than they would be unconstrained.
• Confidence: high that the product exists and works as described. The speed and cost multiples are TypeSafe’s own measurements.

1 to 9 October: open weights arrive

• What happened: Cloudflare’s Workers AI team released its first self-trained models, Clef (27B parameters) and Clef-flash (9B), with weights on Hugging Face under Apache 2.0. A request carries up to 64 questions, and each returns a probability for every allowed answer. Cloudflare describes Clef as being “in the same family as Typesafe’s Jev” and says an existing Jev integration switches by changing the endpoint and model name (Cloudflare changelog). On 9 October it added Clef-omni, built on Qwen3-Omni-30B-A3B-Instruct, which accepts audio and video as well as text and images (Cloudflare blog). On 6 October Perplexity released pplx-decider-v1.1-27b as open weights, at $0.02 per million input tokens, half the price of v1 (Perplexity).
• Why it matters: open weights and a compatible API within three weeks of Jev’s launch mean the format does not belong to one vendor. In Cloudflare’s measurement, Clef-flash answered in a median 38.8 ms, against 524.1 ms for Jev.
• Confidence: high for the release facts. Latency and benchmark leads are Cloudflare’s numbers. Perplexity’s post does not name a license or a base model.

6 October: OpenAI adds a Decisions API

• What happened: OpenAI released the Decisions API in beta on GPT-6 Luna and says it returns typed answers 10x faster than its Responses API (OpenAI changelog). It supports three question types: predicate (the probability that a condition holds), choice (one of your options, with a probability per option) and score (a position on an ordered scale). Uncached input costs $0.10 per million tokens, and output is not charged (OpenAI guide).
• Since then: on 22 October the API left beta and became generally available on GPT-6 Luna. A request now takes up to 400 questions, up from 200 in the beta, and shared input can be cached across requests (OpenAI changelog). Cache writes are charged at GPT-6 Luna’s standard caching rates, and cache hits are discounted (OpenAI guide).
• Why it matters: a frontier lab now sells a decision endpoint on the same model it sells for generation. The guide also draws the boundary. Extraction into your own schema stays with Structured Outputs, and tool requests stay with function calling.
• Confidence: high. Both claims come from OpenAI’s changelog and documentation.

7 and 15 October: Liquid AI moves decision models to the edge

• What happened: Liquid AI released two open-weight models, d1-3B and d1-omni-600M, the second taking text plus image or audio. Its post says d1 models “don’t produce tokens” and answer in a single forward pass. For one question, d1-3B measured between 16 ms (Jetson AGX Thor) and 50 ms (Jetson Orin Nano) on the edge devices tested (Liquid AI). On 15 October Liquid added d1-ultra, an API model it trained for better-calibrated probabilities after finding most decision models overconfident (Liquid AI).
• Why it matters: answers in tens of milliseconds on a Jetson board put decision models inside perception and control loops. Liquid’s demos run d1-3B once per camera frame, and one navigates an environment in NVIDIA Isaac Sim.
• Confidence: high for the release. The latency figures are Liquid’s.

9 October: Microsoft publishes robustness numbers

• What happened: Microsoft released Microsoft-Decision-1, a post-trained Qwen3.5-9B that returns a calibrated probability for each option, in Microsoft Foundry and on OpenRouter at $0.042 per million input tokens. Microsoft reports the highest accuracy in a 36-benchmark comparison of nearly 150,000 questions held out from training. It says the model changed its answer on 1.3% of perturbed inputs on average, and never when options were paraphrased or reordered (Microsoft).
• Why it matters: language models are known to favour certain option positions in multiple-choice questions (Zheng et al., 2023), and a router or classifier cannot afford that. The post also states the calibration target plainly: “a 90% prediction should be right about nine times out of 10.”
• Confidence: medium. Every number comes from Microsoft’s own comparison.

13 October: an independent index

• What happened: Artificial Analysis launched Decision Index v1.0. It combines four evaluations weighted equally: knowledge and hallucination (a private set of 500 questions), reasoning (multiple-choice questions from Humanity’s Last Exam), tool-call decisions (NVIDIA’s When2Call) and safety classification (Ai2’s WildGuardTest). At launch Jev 1.13 led the decision models with 58, ahead of GLiDE and Microsoft-Decision-1 at 54. Two language models run at low effort scored 61 (Claude Haiku 5.5) and 55 (GPT-6 Luna) (Artificial Analysis).
• Why it matters: decision models cost $17 to $199 per million tasks at median times of 0.06 to 0.5 seconds. Claude Haiku 5.5 cost $184 and took 1.4 seconds; GPT-6 Luna cost $131 and took 1.5 seconds. Calibration falls short of the vendors’ claims: “across almost all probability levels, accuracy falls below that probability for all decision models.” Artificial Analysis traces this mainly to the knowledge and reasoning sets. Most decision models were relatively well calibrated on the tool-call and safety sets.
• Confidence: high for the measurements. This is a launch snapshot, and the live leaderboard will move.

What to watch next

• Calibration against language models. Artificial Analysis lists language-model calibration data as “coming soon” (Artificial Analysis). That comparison will show whether decision models are better calibrated or only cheaper.
• A shared interface. Cloudflare accepts Jev requests with an endpoint change, and Cloudflare and Liquid both name their yes/no type noul (Cloudflare, Liquid AI). OpenAI uses the same three types and calls the yes/no type predicate (OpenAI guide). Speculation: a de facto request format forms around these three types before the end of the year.
• Price. TypeSafe, Liquid’s d1-ultra and Microsoft all charge $0.042 per million input tokens, and TypeSafe says its own price may be subsidized (TypeSafe, Liquid AI, Microsoft).
• Agent stacks. Microsoft lists agent controls, model routing and AI judging among its use cases (Microsoft), and OpenAI added a native Decisions integration to GPT-Live, its voice model (OpenAI), on 15 October (OpenAI changelog). My read: the first large deployments will be routing and guardrail checks inside agents, where the answer set already exists.