OpenAI's Decisions API is now generally available, while Liquid AI and Strands ship rival decision models. They return probabilities instead of text. The vendor numbers are unaudited, but the architecture question is real.
A new class of model is reaching production APIs: the decision model. It does not write text. It reads an input and a set of questions, and returns typed answers with probabilities. In October 2026, Strands and Liquid AI (twice) made this concrete. Separately, OpenAI's documentation now lists its Decisions API as generally available. All of them come with claims you should treat as claims.
What shipped
OpenAI's documentation says the Decisions API is now generally available after a public beta. It evaluates text, images and JSON, and OpenAI says it returns typed answers about 10x faster than the Responses API. Requests go to POST /v1/decisions, and gpt-6-luna is the only supported model for now.
There are three question types:
- predicate: the probability, from 0 to 1, that a condition is true.
- choice: one option from a set you supply.
- score: a rating against ordered levels, computed as the probability-weighted average of the level indices, so it can fall between levels.
Since the beta, OpenAI lists prompt caching, JSON input, hosted image URLs, conversation history (including function calls and their outputs), organization usage reporting and Private Link. The limit rose from 200 to 400 questions per request. OpenAI also lists a native integration with GPT-Live.
The documentation draws a boundary. Use Decisions when you need one of these answer types. Use Structured Outputs when you need to generate an object that follows your own JSON schema, or function calling when the model must request a tool call. Decisions returns answers; your application runs whatever action those answers select.
The same idea elsewhere
On 1 October 2026, the Strands team released Strands Decider 2B, an open source decision model. It describes a 2 billion parameter model that can run on a local CPU or GPU. The team reports a median of around 115 ms on widely available hardware, with the published latency graph measured on an RTX 3090 against v18 of the model, and around 153 ms on an M3 MacBook for small tasks. In its own measurement, Strands says the model ranked 3rd of 33 in the 2B class on accuracy and calibration (using the Brier score for calibration) on JevBench's public set. It published the weights, training data and scripts.
On 5 October, Liquid AI introduced d1, with image support. Liquid AI says d1 matches or beats GPT-6.1 Sol on four of six applications it tested and costs 19x to 200x less than GPT-6.1 Sol and Claude Opus 5.5. Its stated price is $0.04 per million input tokens, billed on input only. On 15 October it released d1-ultra at $0.042 per million input tokens. Liquid AI says it compared d1-ultra with d1 and Jev 1.14 across 112 evaluations and says it outperforms Jev 1.14 across all tasks. These are Liquid AI's own benchmarks, run by Liquid AI. In the d1 post (not the d1-ultra post), the methodology note says the company wrote six of the 15 code questions and two of the four compaction sessions after its pipeline was set.
Why the probability matters
The Strands post makes the point most plainly. Decision models, it says, give each decision a reliability score, which it says is not available through frontier LLM inference APIs. Liquid AI makes a related claim: most decision models are overconfident, so it trained d1-ultra for better-calibrated probabilities.
For an engineering lead, that is the useful part. A probability is something you can put a threshold on, route on, and audit.
Strands gives an example of where this could sit. It describes a check that runs before an agent's tool call, asking whether the arguments are grounded in what the user said. The team calls it an illustration rather than a recommendation, with questions and thresholds picked by hand. Its argument is that a decision this cheap can sit in a path where an LLM call never could.
The limits are stated by the vendors
Strands is blunt: this approach is significantly worse at complex problems than reasoning models, and the lack of text generation makes it unsuited to coding, chatbots and summarization. OpenAI's own guidance carves out schema-bound generation and tool calls. Liquid AI, for its part, says decision models can replace expensive calls to language models when the answer is a structured decision.
Strands says it is seeing hybrid setups: an LLM for the hardest decisions, a decision model for the rote ones.
Separately, use cases named across the sources include:
- classifying and routing support tickets,
- filtering content,
- visual inspection of parts or photos,
- policy classification and guardrails,
- context management in agents.
What the sources do not tell you
There is no independent comparison across OpenAI, Liquid AI and Strands on your data. The 10x speedup and the cost multiples are vendor figures. The two deployment shapes also differ: a hosted API with Private Link on one side, open weights you can run locally (Strands) on the other. Liquid AI's models are served through the Liquid AI API (and, per the d1 post, also via Vercel and OpenRouter, with text only for d1 for now). Liquid AI plans to release open weights for upcoming models on Hugging Face soon. For regulated workloads, where data may not leave your environment, that difference may matter more than any benchmark.
What this means for your roadmap
Pick one high-volume classification step in your stack, such as ticket routing or a pre-tool-call guardrail, and run it against two decision models and your current LLM on your own labeled data. Compare accuracy, calibration of the probabilities and cost per thousand decisions, not vendor headlines. Decide early whether data residency requires open weights or Private Link, because that choice narrows the field before quality does.
Sources
- developers.openai.com Decisions | OpenAI API
- liquid.ai Introducing d1: The most capable decision model, now with vision | Blog · 5 अक्टू॰ 2026
- liquid.ai d1-ultra: Introducing the World’s Best Decision Model | Blog · 15 अक्टू॰ 2026
- strandsagents.com Introducing Strands Decider 2B: a small, open source, decision model · 1 अक्टू॰ 2026
This briefing was researched and written by SKYLOGIQ's editorial pipeline: an AI writer and an AI fact-checking editor, with a human review queue for anything the editor refuses. Sources are listed above. How the Briefing is made