Decision Models Compared: Cloudflare Clef, Perplexity pplx-decider and AWS Strands Decider
Three vendors shipped open decision models on the same day. They pick from options instead of writing text, and they differ in size, speed and price.
Jump to 9 sections
A decision model picks from options you define instead of writing text. You give it a situation and a list of questions, and it returns a probability for each allowed answer in one pass. Cloudflare, Perplexity and AWS all shipped open versions on October 1, 2026, and each suits a different job.
This comparison is for developers who build agents and need fast, cheap routing, triage or guardrail checks. Model Drop tracks new models, platforms and tools, and this piece compares the three launches on size, speed, license and fit. We also cover where decision models fall short and how to test them on your own data.
What Is a Decision Model?
A decision model is a classifier built for agents. InfoQ describes decision models as models that pick from predefined options instead of generating text. They classify inputs and return typed outcomes with probabilities, so an agent can route a support request, escalate it or defer to a human.
The output needs no parsing. A normal chat model might answer in a sentence, and your code must then guess what it meant. A decision model returns a number for each allowed option, so your code can act on it directly.
That makes them fast and cheap. One forward pass replaces a long generation, which cuts latency and cost. The trade-off is range, since they cannot explain, summarize or write.
How Do the Three Launches Compare?
All three arrived on October 1 and all three are open-weight. Cloudflare's Clef family is the largest at 27 billion and 9 billion parameters. Perplexity's model is 27 billion. AWS's Strands Decider is about 2 billion.
| Model | Size | License | Where it runs | Reported speed |
|---|---|---|---|---|
| Cloudflare Clef | 27B | Open weights, Apache 2.0 | Workers AI, Hugging Face | 209.3 ms median |
| Cloudflare Clef-flash | 9B | Open weights, Apache 2.0 | Workers AI, Hugging Face | 38.8 ms median |
| Perplexity pplx-decider-v1-27b | 27B | Apache 2.0 | Decisions API, Hugging Face | Under 2 seconds for short input |
| AWS Strands Decider 2B | 1.9B | Apache 2.0 | Local only | 115 ms median on RTX 3090 |
The table above puts the main specs side by side. The speed figures come from each vendor's own testing on different hardware, so do not rank them as if they were run on one machine.
What Does Cloudflare Clef Offer?
Clef is the most capable of the three on inputs. InfoQ reports it accepts text, JSON, images or video, and returns a probability for each allowed option given a schema of typed questions. It has a 64,000-token context window, and Cloudflare's team wrote that it has a vision encoder so it can classify visual content.
Speed is the selling point for Clef-flash. Cloudflare reports a median of 38.8 milliseconds, against 209.3 for the 27B model. Both run on Cloudflare's edge GPUs, which the authors say gives low network latency.
Fine-tuning is available, with limits. Cloudflare offers a service to adapt the models on your own labeled data, such as support triage or bot detection. It starts as a hands-on engagement with Cloudflare engineers, and a self-serve platform has no firm date.
Pair it with a generator when you need action. Cloudflare suggests putting Clef in an agent's hot path and combining it with one of its language models on Workers AI to carry out the decision.
What Does Perplexity pplx-decider Offer?
Perplexity's pitch is a hosted API with a long input window. AI Weekly reports that the Decisions API runs on pplx-decider-v1-27b, an open-weight model fine-tuned from Qwen3.8-27B. It accepts up to 262,144 input tokens and answers yes/no, multiple-choice or ordered-score questions.
Price is low. AI Weekly lists $0.04 per million input tokens with free output. Latency depends on length, with under two seconds for a few hundred input tokens and up to about 23 seconds near the input limit.
The benchmark claim needs caution. Perplexity reports 85.71% across an 11-test panel of 7,210 samples, against 84.51% for TypeSafe's Jev. The same article notes that Jev leads on four of the 11 tests, including WinoGrande and BBH, and that no independent rerun has confirmed the numbers.
What Does AWS Strands Decider 2B Offer?
Strands is the small, local option. eesel.ai's review of Strands Decider 2B says it is 1.9 billion parameters, built on a Qwen3.5-2B-Base torso with a small LoRA adapter. The text-generation head is removed and replaced by a pointer head of about one million parameters that scores each option.
It is fully open. The license is Apache-2.0 and covers weights, code, training data and recipe. It runs on CUDA, Apple MPS or MLX, and CPU, with a 4,096-token default context. There is no hosted API or Bedrock listing, so you run it yourself.
Performance is modest by design. The review reports 0.762 on the public JevBench set of 231 tasks for one recipe, and a median of 115 milliseconds on an RTX 3090. On CPU, short inputs take about 1.6 seconds, and a prompt near 3,000 tokens takes 16 to 18 seconds.
The review also flags weaknesses. It is much weaker than reasoning models on complex problems, and one user reported a bias toward the first option in a list. Note that eesel.ai sells a competing helpdesk product, so weigh its framing with care.
Which One Should You Pick?
Match the model to your constraint. If you need images or video in the decision, Clef is the only one of the three built for it. If you want a managed API and long documents, Perplexity's Decisions API fits. If data must stay in-house or you want a tiny model on your own hardware, Strands is the pick.
Think about volume too. A small local model can handle millions of cheap routing decisions for the cost of a GPU. A hosted API charges per token but removes the work of running a server.
Cost deserves a quick worked example. Say your support bot makes two million routing decisions a month with 400 input tokens each. That is 800 million input tokens, or about $32 at $0.04 per million, before any output. A local model has no per-token bill but needs a GPU and someone to run it.
Microsoft has also entered the space. The Capital and Compute October release tracker lists Microsoft-Decision-1, a decision model post-trained from Qwen3.5-9B on Microsoft Foundry, released October 9. Treat the field as crowded and moving.
For a broader view of how teams route traffic between models, see our guide to LLM gateways and model routing. If you are weighing size and hardware, our piece on small models that run on one GPU covers the cost side.
How Should You Test a Decision Model Before Using It?
Public benchmarks are a weak guide here. A commenter quoted by InfoQ noted that public benchmarks are easy to game, and a Reddit user said a model's ability to know when to defer matters more than raw accuracy. Both points are right.
Build a test set from your own data. Use these steps.
- Collect 200 to 500 real examples with correct labels from your team.
- Include the costly mistakes, such as a refund wrongly approved.
- Run each model and record accuracy and confidence scores.
- Test option order by shuffling the list and checking for bias.
- Set a confidence threshold below which the case goes to a human.
The steps above focus on costly errors and deferral, which matter more than average accuracy. They also reveal data shifts, when your real inputs differ from the training mix.
Keep a fallback in place. When a decision model is unsure, route to a larger model or a person. Our recap of September 2026 model releases shows how fast the larger models are improving, which makes that fallback stronger over time.
Where Do Decision Models Fall Short?
They are not general-purpose. They cannot write code, hold a conversation or summarize a document. The Strands review calls them weak on long, multi-step documents.
They can also be overconfident outside their training mix. A support classifier trained on English tickets may mislead on a new product line or language. Monitoring and periodic retesting are part of the cost of running one.
There is a hype risk too. A Hacker News commenter quoted by InfoQ argued that decision models are not a new paradigm but a low-hanging fruit that has been lying around for years. That may be fair. The news is not the idea but that three vendors made good versions free and open.
At Model Drop, we see these as plumbing, not headlines. Their value shows up in lower bills and faster agents, and the only way to know is to measure it on your own traffic.
Conclusion
Decision models give agent builders a cheaper, faster way to make routing and triage calls. Cloudflare Clef, Perplexity pplx-decider and AWS Strands Decider 2B all launched on the same day, and each suits a different constraint.
Your next step is to pick one candidate, build a labeled test set from your own traffic and measure accuracy, deferral behavior and latency. Keep a larger model as a fallback until the numbers hold up.
