Decision Models Explained: Cloudflare Clef vs AWS Strands Decider (2026)
A factual look at decision models, the category that scores options instead of writing text. Covers Cloudflare Clef at $0.038/M tokens and 38.8 ms median latency, plus AWS's 2B model that runs locally.
The Model That Refuses to Write a Single Sentence
In late September 2026, a model called Jev from TypeSafe AI went viral for something unusual: it does not generate text. You pass it a state and a set of typed questions, and it returns a probability for every allowed answer. No prose, no parsing, no reasoning tokens to wait for.
That single launch has now spawned an entire product category. On October 1, 2026, ten models shipped from seven different labs, and every one of them was a decision model. None came from OpenAI, Anthropic, Google, xAI, Meta, or DeepSeek. For one day, at least, the frontier belonged to newcomers.
Cloudflare's entry, Clef, is the one worth studying in detail, because it ships with concrete numbers and open weights.
What a Decision Model Actually Is
A conventional LLM produces text, and your code parses that text to extract a decision. A decision model collapses those two steps. It reads an input state — text, JSON, images, or video — and scores every option you define, returning calibrated probabilities your application thresholds directly.
Cloudflare's framing is useful: these models are "System One" models in the same family as Jev. They produce bounded structured output cheaply, quickly, and consistently, and they generalize across new categories without retraining. That last property is what separates them from a traditional classifier. You do not need a labeled dataset and a training run to add a new routing category; you add an option to the schema.
The tradeoff is explicit. Because the model scores all options in a single parallel pass, it is far worse at solving complex problems than a reasoning model, and because it cannot emit text, it is unsuitable for coding, chat, or summarization. It is not an LLM replacement. It is a replacement for the classification call you were about to route through an LLM anyway.
The Three Numbers That Matter
Latency: 38.8 ms median
Cloudflare ran 43 evaluation benchmarks and reported median latency of 209.3 ms for Clef and 38.8 ms for Clef-flash, against 524.1 ms for Jev. That makes Clef roughly 2.5x faster than Jev at the median and Clef-flash roughly 13x faster. The p95 figures are 238.6 ms, 122.4 ms, and 536.0 ms respectively.
On October 9, Cloudflare optimized how Clef is served and reported up to a 2x improvement: at roughly 3,400 input tokens, median latency dropped from 616 ms to 305 ms. Model weights were unchanged — the speedup came from the serving stack, including a move to SGLang.
Price: $0.038 per million input tokens
Clef-flash was cut from $0.090 to $0.038 per million input tokens on October 9, which puts it below Jev. Output tokens are not billed, because there are none.
There is a real tradeoff hiding in that price cut: the hosted Clef-flash context window shrank from 64K to 24K tokens. Cloudflare says only 0.24% of requests exceed 24K input tokens. If yours do, use Clef at $0.240 per million with a full 64K window. Self-hosting the Clef-flash weights preserves a 256K context.
The price ladder as of October 9, 2026:
| Model | Price (per M input tokens) | Context |
@cf/cloudflare/clef-flash | $0.038 | 24K |
@cf/cloudflare/clef-omni | $0.150 | 64K |
@cf/cloudflare/clef | $0.240 | 64K |
Accuracy: 94.20 macro-F1 on BANKING77
On BANKING77, a banking intent classification benchmark, Clef scores 94.20 macro-F1, Clef-flash 90.93, and Jev 79.74. On CLINC150+OOS, Clef leads at 97.43 against Jev's 89.27.
But the sweep is not clean, and this is the part most launch coverage skips. Jev still wins When2Call (80.97 vs. 72.37) and BRIGHT (47.52 vs. 45.91). On PhishNChips, the DiffusionGemma variant of Jev leads at 85.35 while Clef sits at 79.60. Pick benchmarks that match your actual workload, not the headline table.
Asking Up to 64 Typed Questions in One Request
Clef supports three question types, and they cover most decision shapes you will encounter:
- noul — returns the probability that a condition holds. Gate an agent's tool call on it.
- choice — picks one option from a set you define, returning the chosen option, a probability per option, and a confidence value. Ticket routing lives here.
- score — rates against an ordered rubric, returning a probability-weighted score plus a probability per level. Severity and priority scoring live here.
A concrete request: hand it the state "Checkout has been failing for every customer for the last hour" and ask three questions at once — is this urgent (noul), which team owns it (choice across billing, technical, sales), and how severe is the impact (score across No impact, Minor, Major, Critical). Your application thresholds the returned numbers and acts.
AWS Shipped a 2B Model That Runs on Your Laptop
The same week, AWS's Strands Labs released Strands Decider 2B. Architecturally it is clean: take a pre-trained Qwen3.5-2B torso, remove the language model head so it can no longer generate text, and replace it with a pointer head that scores each candidate answer. The added head is just over a million parameters, and the torso carries a rank-16 LoRA adapter. Weights, training data, and scripts are all public.
Because it is 2 billion parameters, it runs locally. AWS reports a median of about 115 ms on an RTX 3090 and about 153 ms for small tasks on an M3 MacBook, with latency scaling roughly linearly in task size.
The example in AWS's own post is the most instructive part. An agent with an eager system prompt is about to guess a city and call a weather tool the user never specified. Before the call executes, Strands Decider answers two yes/no questions: are the tool's argument values grounded in something the user actually said, and is it too early to call this tool. A few lines of Python turn those probabilities into a Proceed, Deny, Confirm, or Guide action. AWS is explicit that this is an illustration, not a recommendation — the questions, threshold, and policy were hand-picked.
Where This Goes
Four patterns are already visible in production.
That last one is probably the largest practical win, and it is the one AWS's post pushes hardest: a decision this cheap can sit in a request path where an LLM call never could.
Getting Started
Clef is available on Workers AI via the env.AI.run() binding or the REST API at /ai/run, and it works through AI Gateway. It follows the same System One API as Jev, so switching an existing integration means changing the endpoint and the model ID.
For Strands Decider, the fastest path is pip install strands-decider, then a single CLI call with a --state and a --choice.
Start hosted, measure accuracy against your own data, and only move to self-hosting — where the context windows are far larger — once latency or cost actually becomes the binding constraint.