OpenAI, Microsoft, and Cloudflare all launched one this month — here's what they actually are
In the first week of October 2026, something unusual happened in AI: four major companies launched the same new product category within days of each other. OpenAI released its Decisions API (October 6), Cloudflare shipped open-weight Clef models (October 1, then cut prices October 9), Microsoft unveiled Decision-1 (October 9), and Liquid AI open-sourced its d1 decision models (October 7). This guide explains what decision models are, how they differ from ChatGPT and Claude, and whether you should care — honestly, with no hype.
📰 What Happened This Month
- Oct 1: Cloudflare releases Clef (27B) and Clef-flash (9B), open-weight under Apache 2.0
- Oct 5–7: Liquid AI releases d1-3B and d1-omni-600M, open decision models that run locally
- Oct 6: OpenAI launches the Decisions API in public beta, powered by gpt-6-luna
- Oct 8: llama.cpp adds a decision-model endpoint, so these models run on your own machine
- Oct 9: Microsoft launches Decision-1 at $0.042 per million tokens; Cloudflare cuts Clef-flash to $0.038 and adds audio/video with Clef-omni
No consumer blog has explained this category yet. This is that explainer.
A decision model is an AI model that does exactly one job: given a situation and a fixed list of possible answers, it returns a probability for each answer. It never writes a sentence. It never chats. It never explains itself.
Here's a concrete example. A customer writes to your support inbox: "My package arrived broken and I want my money back." A traditional approach sends that message to a large language model like GPT or Claude and asks, "Which department should handle this?" The model writes back a paragraph, and your code has to parse the paragraph to figure out the answer — hoping it picked one of your real departments and didn't invent a new one.
A decision model skips all of that. You give it the message plus your fixed options — Billing, Shipping, Technical Support, Other — and it returns something like:
Billing: 0.08 | Shipping: 0.87 | Technical Support: 0.03 | Other: 0.02
→ Decision: Shipping (confidence 0.87)
Your software reads those numbers directly and routes the ticket. No parsing, no invented categories, no wasted output tokens. The category gets its nickname — "System One" models — from psychologist Daniel Kahneman's idea of fast, intuitive System 1 thinking versus slow, deliberate System 2 reasoning. ChatGPT is System 2. Decision models are System 1.
Three Question Types
Nearly every decision model on the market supports the same three question shapes:
- Yes/No (predicate): "Is this message spam?" → returns a single probability, e.g. 0.94.
- Choice: "Which category fits best?" → returns one winner plus a probability for every option.
- Score: "Rate this review from 1 to 5" → returns a probability-weighted position on your ordered scale.
Multiple questions can share one request — for example, asking "is this urgent?", "which team owns it?", and "rate the customer's tone 1–5" all in a single call over the same message.
This is the comparison that matters most, because the two categories look similar from the outside but are built for opposite jobs:
| Decision Models | ChatGPT / Claude (LLMs) | |
|---|---|---|
| Output | Numbers: probabilities, choices, scores | Text: sentences, paragraphs, code |
| Speed | Milliseconds (8–200 ms typical) | Seconds (1–10+ s typical) |
| Cost model | Input tokens only; output is free | Input + output tokens billed |
| Reliability | Can only pick from your options | Can invent answers you never listed |
| Best for | Routing, classification, moderation, agent control | Writing, coding, conversation, reasoning |
| Can it chat? | No — not designed for it | Yes — that's the point |
The honest way to think about it: decision models don't replace ChatGPT or Claude — they sit next to them. In a modern AI agent, the LLM does the thinking and writing, while a decision model handles the rapid-fire judgment calls around it: should the agent continue, retry, escalate to a human, or call a different tool? If you're choosing between the two chatbots themselves, see our ChatGPT vs Claude comparison.
OpenAI Decisions API 🔥 BIGGEST NAME
What: A dedicated API endpoint (POST /v1/decisions) that returns typed answers instead of text
Model: gpt-6-luna (currently the only supported model)
Status: Public beta since October 6, 2026; general availability expected "in the coming weeks"
Price: $0.10 per 1M input tokens; output tokens free
Input: Text and images (images must be inline base64 — hosted URLs are refused)
OpenAI says answers come back about 10x faster than asking the same question through its Responses API. That speed claim comes from OpenAI itself — no independent benchmark has confirmed it yet.
Microsoft Decision-1 🔥 CHEAPEST BIG-LAB OPTION
What: A decision-scoring model for routing, classification, verification, and agent control
Model: Post-trained from Qwen3.5-9B (open-weight base)
Status: Generally available via Microsoft Foundry
Price: $0.042 per 1M input tokens; output tokens free
Input: Text only — no images, audio, or video
Microsoft claims 35x faster median latency than GPT-6 Sol — vendor-reported, not independently verified. The model card is explicit about limits: no text generation, no explanations, and not for sole automated decisions on credit, employment, housing, healthcare, or legal rights.
Cloudflare Clef Family OPEN WEIGHT
What: Open-weight decision models (Apache 2.0) on Workers AI and Hugging Face
Models: Clef (27B, multimodal), Clef-flash (9B, latency-optimized), Clef-omni (30B MoE, adds audio + video — launched Oct 9)
Status: Available now; weights downloadable
Price: Clef-flash cut to $0.038 per 1M tokens on Oct 9; Clef at $0.24
Input: Text, JSON, images — and audio/video on Clef-omni
Because the weights are open, you can self-host or fine-tune on your own data. The Oct 9 price cut made Clef-flash cheaper than competitors — deliberate competitive positioning.
Liquid AI d1-3B & d1-omni-600M RUNS ON YOUR LAPTOP
What: Tiny open decision models for edge and local deployment
Models: d1-3B (3B params, text + images), d1-omni-600M (600M params, text + image or audio)
Status: Released Oct 5–7 on Hugging Face; day-one llama.cpp support
Price: Free to run locally; hosted tier exists
Input: Text, images (d1-3B); plus short audio clips (d1-omni-600M)
The headline number: 8 milliseconds per question on an RTX 4090. The catch is the license: free commercial use only for companies earning under $10M/year.
TypeSafe Jev THE INCUMBENT
What: The decision model that started the wave; popularized via Vercel's AI Gateway in September 2026
Status: Closed weights, API-only
Price: $0.042 per 1M input tokens
Input: Text and JSON
Jev defined the category's API shape — Cloudflare made Clef API-compatible with it, so migration is mostly changing an endpoint.
Pricing Comparison
| Model / API | Input price (per 1M tokens) | Output price | Weights |
|---|---|---|---|
| Cloudflare Clef-flash | $0.038 | Free | Open (Apache 2.0) |
| Microsoft Decision-1 | $0.042 | Free | Closed |
| TypeSafe Jev | $0.042 | Free | Closed |
| Cloudflare Clef (27B) | $0.24 | Free | Open (Apache 2.0) |
| OpenAI Decisions API | $0.10 | Free | Closed |
| Liquid AI d1-3B | Free (self-hosted) | Free | Open (revenue-capped) |
Prices are per vendor announcements as of October 9–11, 2026 — treat them as a snapshot, not a guarantee. The pattern is clear: every vendor charges for input only, because decision models produce no output tokens to bill.
- Support ticket routing: classify and route thousands of incoming messages per hour without an LLM writing a paragraph per ticket.
- Content moderation: score user content against safety policies in milliseconds.
- AI agent control: the "traffic cop" inside an AI agent — deciding whether to continue, retry, escalate to a human, or switch tools at each step.
- LLM-as-a-judge: scoring AI outputs on a rubric far cheaper than asking a big model to write an evaluation.
- Fraud and spam filtering: yes/no probability scores on transactions or messages at high throughput.
- Visual inspection: "is the screen cracked?" from a customer photo — a yes/no answer with calibrated confidence, no chat needed.
Who Should Care — and Who Shouldn't
You should care if you build software that makes repetitive classification decisions at scale, you're building AI agents and want a cheap control layer, or you're already paying an LLM to output JSON you then parse.
You can safely ignore this if you're a casual ChatGPT or Claude user. Decision models are developer infrastructure, not consumer products. If you just want a better chatbot, our ChatGPT vs Claude guide is the one that matters to you.
⚠️ Honest Limitations
- Speed claims (10x, 35x) are vendor-reported and not yet independently verified.
- Decision models give no explanations — you get a number, not a reason.
- The category is weeks old. APIs are in beta, pricing may change.
- Accuracy is task-specific. Test on your own data before trusting.
Frequently Asked Questions
What is a decision model in simple terms?
An AI model that answers fixed multiple-choice-style questions with probabilities instead of writing text. You give it a situation and a list of allowed answers; it tells you which answer is most likely and how confident it is — in milliseconds, with no chat involved.
Are decision models replacing ChatGPT?
No. They do a different job. ChatGPT and Claude generate text, write code, and hold conversations. Decision models make rapid structured judgments (routing, scoring, classifying). In practice they work together: the LLM thinks, the decision model decides.
Which decision model is cheapest?
As of October 2026, Cloudflare's Clef-flash is cheapest among hosted APIs at $0.038 per million input tokens, followed by Microsoft Decision-1 and TypeSafe Jev at $0.042. OpenAI's Decisions API is $0.10. Liquid AI's d1-3B is free if you run it yourself. Prices change fast — verify before building.
Can I run a decision model on my own computer?
Yes. Liquid AI's d1-3B and d1-omni-600M, Cloudflare's Clef models, and several smaller models have open weights and run in llama.cpp (builds from October 8, 2026 or later). A gaming GPU handles them comfortably.
Do decision models work with images and video?
Some do. OpenAI's Decisions API and Cloudflare's Clef models accept images; Cloudflare's Clef-omni (launched Oct 9) also handles audio and video. Microsoft's Decision-1 is text-only.
Decision models are the first genuinely new AI product category since chatbots — and this time, the infrastructure arrived all at once. They're not for chatting. They're for the millions of tiny judgments inside software that used to require either slow LLMs or hand-written rules.
The category is weeks old, the benchmarks are vendor-supplied, and the pricing is a land-grab. But the underlying idea — probabilities instead of paragraphs — is sound, and it's not going away.
Related: What Are AI Agents? Complete Beginner's Guide | ChatGPT vs Claude: Which Is Better?
Last updated: October 11, 2026. Pricing and availability change quickly — verify current terms on each vendor's site before building.
No comments:
Post a Comment