Cloudflare Clef alternatives: five decision models compared

Which decision model should you use?

Five models, five reasons to investigate one. Start with your constraint, then check the evidence and limits below. Clef, Clef-flash and Jev have recorded ClefBench measurements. Perplexity and Strands are guide only — not tested here.

1. Highest yes/no accuracy in our set; video-capable weights

   Clef

   $0.240/M input · 638 ms median · 98.33% yes/no · 96.67% choice

2. Lowest current direct API price among our tested models

   Clef-flash

   $0.038/M input · 436 ms median · 96.67% yes/no · 96.67% choice

3. Text only; exact category choices and shortest wait in our set

   Jev 1.13

   $0.042/M input · 356.5 ms median · 96.67% yes/no · 100.00% choice

4. Very long states or more than 64 questions

   Perplexity Decider v1.1

   $0.020/M input · Guide only — not tested here

5. Operate a smaller open model yourself

   Strands Decider v19

   Self-hosted; budget for hardware and operations · Guide only — not tested here

Cloudflare also lists Clef-omni at $0.150 per million input tokens in its official pricing, checked October 10, 2026. It is a pricing-only reference here: we have not tested it or verified its full specification, so it is outside the five-model comparison.

The five models, field by field

Use the licence and self-hosting columns to narrow deployment choices, then inspect measured waiting time and accuracy. Context and question limits refer to the named providers' interfaces, not this website's playground allowances. A larger context window does not establish better accuracy on long documents.

Which of these can you self-host?

Running downloaded weights and reselling a provider's hosted access are different arrangements. The following list adds Laya and Kev as smaller deployment options; neither has been benchmarked by ClefBench.

Cloudflare's model card includes Transformers inference code and an SGLang serving recipe using Docker. Hugging Face also exposes integration entry points such as vLLM, but a generic chat-serving example is not proof that it runs the custom decision head correctly. Follow model-specific instructions and validate typed outputs.

Size the hardware for your workload. Cloudflare's published inference example uses a single H200; that is a documented setup, not a universal minimum requirement. A 421M Laya checkpoint and a 9B Clef-flash occupy different memory brackets from Clef 27B. Quantization, context length and concurrency change the requirement. We do not promise that any consumer GPU can run a given variant.

The Apache 2.0 licence permits commercial use subject to its conditions, including applicable notices and redistribution requirements. Operating those weights yourself does not grant permission to resell another company's API. For example, TypeSafe's customer agreement, sections 2.1–2.3, distinguishes integration in customer applications from making the service available as a standalone service. Check the terms for the actual arrangement you plan to use.

Official figures vs what we measured

Cloudflare's published comparison gives median latency of 209.3 ms for Clef, 38.8 ms for Clef-flash and 524.1 ms for Jev 1.13. On that benchmark, Clef's median is about 2.5 times lower than Jev's.

On our recorded path, the order reverses:

Our times include the network round trip from a local macOS runner through OpenRouter to the pinned provider. They measure end-to-end waiting on that path, not internal model latency. Different tasks, hosting paths and timing boundaries make the two sets unsuitable for a direct speed-ratio claim.

Clef led on yes/no accuracy in our set. Jev led on exact choices, mean absolute rating error and median wait. All three ran the same 60 short synthetic cases, three times each; the reference answers were written by ClefBench and have not received an independent second review. The historical costs above are recorded usage costs, not a recalculation at today's direct API prices.

Read the full methodology and original data, or inspect support routing, invoice review and risk triage. Which errors matter depends on your workflow; these results do not establish a universal winner.

What it costs at your volume

For the hosted APIs listed below, input tokens drive model usage charges. Compare the same workload: 100,000 decisions × 500 input tokens = 50 million input tokens per model. The example includes context and questions; it excludes free allowances, base plans, account fees and taxes. It does not estimate self-hosting costs.

At the prices checked on October 10, Clef-flash's direct input rate is about 9.5% lower than Jev's. Its context is 24,576 tokens, compared with Clef's 65,536. Jev has two separate budgets: 64k tokens for the entire request and 32k for state plus the longest question. Cheap input tokens do not help if required evidence is truncated.

Change the workload below using the same calculator as our official API pricing page. Its word-to-token conversion is an estimate; actual tokenization, question length and media affect the bill. Clef-omni's presence in the cost estimate does not imply that it is available in our playground.

A ClefBench subscription pays for this website's online features and credits. Direct provider charges and self-hosting infrastructure are separate costs.

What people building with these models report

These are reports from other people's setups, not ClefBench measurements. Sources were checked on 2026-10-10; dates below distinguish an author's publication date from our access date. Star and vote counts are omitted because they change and do not establish quality.

• Local game-agent comparison — Reddit, accessed 2026-10-10. In “Benchmarking decision models is fun — Clef Q8 vs Jev”, the author reports Clef 27B running locally on an RTX 5090 and winning six duels to Jev 1.13's three. The same author reports only one win in 22 matches against a dedicated reinforcement-learning bot. A reply challenges the single-decision setup and points to Jev's ability to evaluate many questions about one state. This is an anecdotal game test, not evidence of general superiority or equivalent throughput.
• Small Workers AI experiments — DEV, published 2026-10-04. 灯里/iku's hands-on report finds simple routing useful but reports missed details in images and multi-intent requests. The author highlights option IDs as model input, sorted object keys, state-tail truncation in the published code and possible interactions between questions. They explicitly do not establish that Workers AI uses every local-code behavior. The article's old Flash price and context are superseded by the current official specifications above.
• Code and deployment tools — GitHub, accessed 2026-10-10. cobanov/awesome-jev collects source-linked projects, while jaredpalmer/kev publishes weights, serving instructions and model-specific hardware notes. Use these repositories to inspect how a deployment works; inclusion in a resource list is not a benchmark endorsement.

Try the interface, then compare your own rules

For direct Cloudflare integration, follow the Clef API tutorial. It includes the authenticated request, model selector and complete question schema. For the model's capabilities, see the full Clef spec.

1. Collect representative examples you are permitted to process, including ambiguous boundaries and costly failures.
2. Write reference answers before running models; ask a second person to review disputed labels.
3. Keep context, question types and allowed answers fixed. A changed rubric is a different test.
4. Inspect false positives, false negatives and variation across repetitions, not just aggregate accuracy.
5. Measure waiting time and billed input in your intended hosting environment.

You can try a browser comparison or save a private evaluation set with a subscription. For narrower reading, use Clef vs Jev, Clef vs Perplexity Decider and Clef vs Strands Decider.

What we have not tested

• Perplexity Decider, Strands Decider, Clef-omni, Laya and Kev have no ClefBench measurements here.
• The measured set has 60 synthetic short cases: 20 support, 20 invoice and 20 risk. It is not a production-traffic sample.
• Reference answers have no independent second review. Accuracy reflects those references.
• Strands v19 is a fixed documentation target, not the project's latest model.
• We have not evaluated long-context accuracy, images, video, self-hosted deployments or production concurrency.
• Waiting time includes our network path. Hardware, geography and service conditions may change yours.