Which decision model API should I use for routing and classification?

As of October 11, 2026: Microsoft-Decision-1 is the cheapest hosted option ($0.042 per million input tokens, output free, on Microsoft Foundry and OpenRouter) and the best fit for text or JSON routing inside Azure. OpenAI's Decisions API ($0.10 per million input tokens on gpt-6-luna, output and caching free, public beta since October 6, 2026) is the pick when decisions must look at images or you already run on OpenAI. Liquid AI's open-weight d1-3B is the pick for on-device, edge or data-residency work: it answers a question in about 8 ms on an RTX 4090 and runs under llama.cpp.

What is a decision model?

A decision model does not write text. You send a state (text, JSON or images) and a set of typed questions, and it returns a probability that a statement is true, a choice from fixed options with a probability for each, or a score on an ordered scale, in a single pass. Software can threshold those probabilities to act, defer or ask a human, which is why decision models are used for routing, triage, moderation, LLM-as-a-judge scoring and agent guardrails.

How much does the OpenAI Decisions API cost?

With gpt-6-luna, the only supported model, input costs $0.10 per million tokens and there are no charges for output tokens, cache reads or cache writes. The API is in public beta (announced October 6, 2026) and OpenAI says it expects general availability in the coming weeks.

Is a decision model better than asking an LLM for JSON?

For fixed-answer questions, usually yes: it is faster, cheaper because there are no output tokens, and returns calibrated probabilities instead of a label you have to trust. Use an LLM with structured outputs when you need to extract fields, write an explanation or produce free-form JSON.

Quick Answer

OpenAI Decisions API vs Microsoft-Decision-1 vs Liquid d1

Published: October 11, 2026

The short answer

Three decision-model options shipped in the same week of October 2026, and they split cleanly by where the decision has to run. As of October 11, 2026:

Microsoft-Decision-1 — cheapest hosted decision API: $0.042 per million input tokens, output free, on Microsoft Foundry and OpenRouter. Text or JSON in.
OpenAI Decisions API — the hosted option that also reads images: $0.10 per million input tokens on gpt-6-luna, no output or cache charges. Public beta since October 6, 2026.
Liquid AI d1-3B — open weights you run yourself: 3.12B parameters, text and images, about 8 ms per decision on an RTX 4090 and 50 ms on a Jetson Orin Nano.

All three answer the same three kinds of question — is this true?, which option?, how much? — and return probabilities your code can threshold.

The comparison

	OpenAI Decisions API	Microsoft-Decision-1	Liquid d1-3B
Released	Public beta October 6, 2026	October 2026 (Foundry docs dated October 7)	Open weights October 7, 2026
Model	`gpt-6-luna` (only option)	Post-trained from Qwen3.5-9B	Built on LFM2.5-VL-3B, 3.12B params
Inputs	Text and images (inline base64 only)	Text or JSON state	Text, JSON and images; 32,768-token context
Question types	`predicate`, `choice`, `score`	`noul`, `choice`, `score`	Yes/no, choice, score
Price	$0.10 / 1M input; output, cache read and write free	$0.042 / 1M input; output free	Free weights; you pay for the hardware
Where it runs	OpenAI API (`POST /v1/decisions`)	Microsoft Foundry (GlobalStandard, DataZoneStandard), OpenRouter	Your GPU, Mac or Jetson; llama.cpp and Transformers
Vendor speed claim	About 10x faster than GPT-6 Luna via Responses	P50 about 35x faster than GPT-6 Sol	8 ms on RTX 4090, 30 ms on Apple M5 Pro
Status	Beta; GA “in the coming weeks”	Generally available in Foundry	Research release; d1-omni-600M experimental

Picks by job

Routing support tickets or agent requests inside Azure: Microsoft-Decision-1. It is less than half the price of OpenAI’s option on input and Microsoft publishes robustness numbers: the decision flips on 1.3% of paraphrased or reordered requests on average, and never when options are shuffled. DataZoneStandard deployments keep inference inside one data zone, which matters for EU data. Microsoft says it will rebase the model on its own MAI and OpenAI models later.

Checking photos or screenshots: OpenAI Decisions API. It is the only hosted option here that takes images — product-damage checks, UI-state checks for computer-use agents, document-type classification from a scan. Two limits: images must be inline base64 data URLs (no hosted URLs or file_id), and decisions that depend on an earlier answer need separate requests.

On-device, edge or no-cloud decisions: Liquid d1-3B. Liquid reports 48.57 on the Decision Index v0.2.1, ahead of every model under 10B and level with the 35B Decider model. It answers three questions over one state in 21 ms on an RTX 4090. The smaller d1-omni-600M adds audio input but scores 15.95 on the same index, so treat it as an experiment. Liquid’s announcement does not name the licence text, so read the model card before shipping commercially.

What decision models are good for

Model and tool routing — pick the cheapest model that can handle a request (see our LLM router and gateway comparison).
Agent guardrails — “does this tool call touch production data?” as a probability, with a threshold that sends borderline calls to a human.
LLM-as-a-judge — score an answer against a rubric without paying for a written critique.
Moderation and prompt-injection screening before content reaches the main model.
Labelling — Microsoft’s Xbox research team sorted 10,000+ pieces of player feedback into fixed themes and reports quality competitive with GPT-6 Sol at more than 14 times the speed.

They are the wrong tool for extracting fields or writing explanations; use structured JSON output from an LLM for that.

Cost at scale

A routing step that reads a 1,500-token request costs about $0.00015 on the OpenAI Decisions API and $0.000063 on Microsoft-Decision-1. At 10 million decisions a month that is roughly $1,500 versus $630, before any LLM call. A self-hosted d1-3B on one RTX 4090 handles about 475 packed states per second by Liquid’s measurement, so a single workstation-class GPU covers most routing volumes. Compare with LLM prices on our current API prices page.

How to choose in three questions

Does the decision need to see an image? Yes → OpenAI Decisions API or d1-3B.
Can the data leave your network? No → d1-3B on your hardware, or Microsoft-Decision-1 with DataZoneStandard.
Is cost per decision the main constraint? Yes → Microsoft-Decision-1 hosted, d1-3B self-hosted.

Java teams can already call OpenAI’s endpoint: LangChain4j 1.22.0 added OpenAiDecisionModel.

Last verified: October 11, 2026.