Clef makes sense inside a Cloudflare Worker and almost nowhere else: Perplexity's Decider scores higher, reads images too and costs a twelfth as much. We rate it 3.6 out of 5.

How much does Clef cost?

Workers AI: $0.24 per 1M input tokens (images count as input, at most 1,024 tokens each); output free. Weights free under Apache 2.0.

What are the downsides of Clef?

$0.24 per 1M input tokens, about 6x Jev and 12x pplx-decider; 68th of 228 on Decision Index 0.4.1, below Jev 1.14; public scores well above private ones on the community board; self-hosting documented only on H200-class GPUs; images must be inlined; remote URLs are rejected.

How much does Clef cost for a team or in production?

10M text decisions a month at 500 input tokens each is 5B tokens: about $1,200 on Workers AI, against roughly $210 on Jev and $100 on pplx-decider at list prices.

Is Clef safe for company data?

Cloudflare says it does not read, store or train on Clef requests or responses unless you opt into fine-tuning; self-hosted Apache 2.0 weights keep data entirely in-house.

What are the best Clef alternatives?

The closest alternatives we have reviewed are Perplexity Decider (pplx-decider), d3 (vLLM Semantic Router), Jev, GLiNER2.5-Decide and Laya.

Clef review: pricing and verdict (2026)

What it is

Clef is a family of decision models Cloudflare released on 1 October 2026, with weights on Hugging Face under Apache 2.0. Instead of writing text, it returns a probability for every allowed option of every question in one forward pass: a small "joint schema head" reads the final state of a frozen Qwen3.8-27B backbone. Clef-flash uses a Qwen3.5-9B backbone for latency-critical work, and Clef-omni, published later, builds on Qwen3-Omni-30B-A3B.

It answers Jev's three question types (choice, score and the yes/no "noul") and accepts Jev request bodies unchanged. Unlike Jev, it reads images and video. On Workers AI it costs $0.24 per million input tokens with output free, takes 1 to 64 questions and up to four embedded images per request, and has a 65,536-token context window.

The model card gives a Python loader and an SGLang server that answers at /v1/systemone; Cloudflare tested it on a single H200 and publishes launch commands for H200, B200 and B300.

Our review

Clef is a capable decision model sold at a price its scores do not justify. Its launch pitch was that it reads images and video, which Jev cannot, and that its weights are open. Both remain true, but Perplexity's pplx-decider-v1.2 also takes images under Apache 2.0, and at $0.02 per million input tokens against Clef's $0.24 it undercuts Cloudflare twelvefold.

Cloudflare's own launch results are strong on narrow operational tasks: 94.2 on BANKING77 against Jev's 79.7, and a 209 ms median latency against Jev's 524 ms, with Clef-flash at 38.8 ms. The same table shows Jev ahead on When2Call (81.0 against 72.4) and on BRIGHT retrieval, and Clef-flash falls to 66.8 on CLINC150 with out-of-scope queries, where Jev scores 89.3. These are vendor figures on vendor-chosen tests.

The independent board is less flattering. On the community Decision Index 0.4.1, updated on 22 October and ranked only on private tests nobody can train on, Clef sits 68th of 228 at 54.92, behind Jev 1.14 at 60.24 and Perplexity's Decider v1.2 in fourth at 67.98. Its public-benchmark score of 61.71 is nearly seven points higher, a gap that points to familiarity with public test sets rather than general skill. Clef-flash ranks 103rd and Clef-omni 101st; on the same site's vision board Clef-omni places 29th of 55, and the 27B Clef has not been entered.

What Cloudflare does offer is placement and terms. The model runs on the same network as Workers, so a Worker classifying an upload makes no extra hop to a third-party API, and Cloudflare says it does not read, store or train on requests or responses unless a customer opts into its new fine-tuning service. That service starts as engineer-led engagements, which makes Clef look more like a base for a custom model than a finished product. Images must also arrive as embedded bytes, since remote URLs are rejected, so a Worker has to fetch and inline every picture before it can ask a question about it.

Where it earns its place

A Workers developer at a marketplace screening seller photo uploads
Flag listing images that show prohibited items before they go live
A Worker inlines each upload and asks one choice and one noul question. An image costs at most 1,024 tokens, so 1M uploads a month is roughly $260 at $0.24 per million; the same tokens on pplx-decider would be about $22, plus a network hop.
A security engineer triaging user-reported phishing at a mid-sized company
Score reported emails and screenshots for phishing likelihood
Cloudflare uses Clef for its own threat intelligence, and on its PhishNChips test Clef scores 79.6 against Jev's 62.6. Each report gets a probability rather than prose, so thresholds can be tuned; check against your own mail first, because the figure is Cloudflare's.
An ML platform engineer with one H200
Self-host a decision model behind an existing Jev integration
Run the SGLang container from the model card and point the Jev client at /v1/systemone. The card's example binds port 30000 on 0.0.0.0 with no authentication, so put it behind a private network or an authenticating proxy first.

Before you commit

Setup: On Workers AI, call @cf/cloudflare/clef with a state and 1 to 64 questions; self-hosting is one SGLang Docker command on an H200, B200 or B300.
Integrations: Jev/System One request and response format, Workers AI bindings, AI Gateway, and SGLang or Transformers 5.10 for self-hosting.
Data and privacy: Cloudflare says it does not read, store or train on Clef requests or responses unless you opt into fine-tuning; self-hosted Apache 2.0 weights keep data entirely in-house.
Cost at scale: 10M text decisions a month at 500 input tokens each is 5B tokens: about $1,200 on Workers AI, against roughly $210 on Jev and $100 on pplx-decider at list prices.
Limits: 65,536-token context including images and questions; up to four PNG, JPEG or WebP images of 4 MiB each, 8 MiB decoded in total and a 13 MiB body; Workers AI rate limits for Clef are not published on the model page.

What's good

Reads up to four images per request; Jev reads none
Apache 2.0 weights plus a hosted Workers AI endpoint
Accepts Jev request bodies unchanged
Clef-flash median latency of 38.8 ms in Cloudflare's tests
Cloudflare says it does not store or train on requests

What's not

$0.24 per 1M input tokens, about 6x Jev and 12x pplx-decider
68th of 228 on Decision Index 0.4.1, below Jev 1.14
Public scores well above private ones on the community board
Self-hosting documented only on H200-class GPUs
Images must be inlined; remote URLs are rejected

Best for

Teams already running on Cloudflare Workers who want typed decisions about images next to their code and value data terms over leaderboard rank.

Alternatives

Perplexity Decider (pplx-decider)

PerplexityDecision modelsOpen weights

4.5/5

Perplexity's open 27B and 4B decision models, also sold as a $0.02-per-million-token API: text and images in, probabilities out.

d3 (vLLM Semantic Router)

vLLM Semantic RouterDecision modelsOpen weights

4.0/5

vLLM Semantic Router's open 27B decision model with video input and stock vLLM serving, already overtaken by its own d3-turbo.

Jev

TypeSafe AIDecision models

3.8/5

TypeSafe AI's hosted decision model: typed choices, scores and yes/no probabilities for $0.042 per million input tokens, output free.

GLiNER2.5-Decide

FastinoDecision modelsOpen weights

3.6/5

Fastino's sub-billion-parameter zero-shot classifier: pass any labels at call time and get one back on a CPU, no prompt or generation.

Laya

Convai InnovationsDecision modelsOpen weights

3.4/5

Apache 2.0 decision model on ModernBERT that answers typed questions in tens of milliseconds, but needs fine-tuning to be accurate.

Clef: common questions

Is Clef worth it?: Clef makes sense inside a Cloudflare Worker and almost nowhere else: Perplexity's Decider scores higher, reads images too and costs a twelfth as much. We rate it 3.6 out of 5.
How much does Clef cost?: Workers AI: $0.24 per 1M input tokens (images count as input, at most 1,024 tokens each); output free. Weights free under Apache 2.0.
What are the downsides of Clef?: $0.24 per 1M input tokens, about 6x Jev and 12x pplx-decider; 68th of 228 on Decision Index 0.4.1, below Jev 1.14; public scores well above private ones on the community board; self-hosting documented only on H200-class GPUs; images must be inlined; remote URLs are rejected.
Who is Clef best for?: Teams already running on Cloudflare Workers who want typed decisions about images next to their code and value data terms over leaderboard rank.
How much does Clef cost for a team or in production?: 10M text decisions a month at 500 input tokens each is 5B tokens: about $1,200 on Workers AI, against roughly $210 on Jev and $100 on pplx-decider at list prices.
Is Clef safe for company data?: Cloudflare says it does not read, store or train on Clef requests or responses unless you opt into fine-tuning; self-hosted Apache 2.0 weights keep data entirely in-house.
What are the best Clef alternatives?: The closest alternatives we have reviewed are Perplexity Decider (pplx-decider), d3 (vLLM Semantic Router), Jev, GLiNER2.5-Decide and Laya.

Answers checked Oct 22, 2026.