Jev vs. Clef: A Practical Comparison of System One Decision Models

Two weeks apart in September and October 2026, TypeSafe AI and Cloudflare released models built around almost the same idea.

TypeSafe AI launched Jev on September 15, describing it as the first “System One” model. Cloudflare followed on October 1 with Clef, designed to be API-compatible with Jev and aimed at the same class of high-volume classification and decision tasks.

Unlike a traditional LLM, neither model generates prose. You send in some content along with a schema of typed questions, and the model returns structured JSON with probability scores for each answer.

The idea borrows from Daniel Kahneman’s Thinking, Fast and Slow. System 1 is fast and intuitive, while System 2 is slower and more deliberate. Jev and Clef are designed to handle the fast, repetitive decisions that production pipelines may need to make thousands of times an hour, leaving a generative LLM to deal with the smaller number of cases that require deeper reasoning or natural-language output.

That makes the comparison unusually direct. These models target the same workload, expose similar APIs, and can be tested on the same data.

This article looks at what Jev and Clef are, how their APIs work, and how they compare in practice using a head-to-head test on 200 YouTube comments, measuring accuracy, speed, and cost.

What System One models actually do

You don't chat with these models. You give them two things: a state (any text, JSON, or in Clef's case, images and video) and a schema of questions. They return typed, probabilistic answers in a single forward pass with no text generation involved.

Three question types are supported by both models:

noul (Yes/No): Returns a probability rather than a binary answer. "Is this spam?" returns {"spam": 0.88}, an 88% confidence that it is. Developers can set their own thresholds rather than trusting a hard cutoff.

choice (Multiple Choice): Defines a set of options and returns the most likely one with a confidence score. "Tone?" with options ["praise", "critique", "joke", "question", "neutral"] returns the winning category.

score (Rating Scale): Returns a float between 0 and 1. "How toxic?" returns 0.22.

The structured output is what makes these models valuable for software. No string parsing, no regex on model output, no "prompt wrangling" to get consistent formatting. The response is directly usable in application logic.

Note: the question type is noul, not nou as some early documentation printed it. The l matters for the API.

The two models

Jev by TypeSafe AI

Jev was built by Diogo Almeida (previously at OpenAI, where he worked on the instruction-following research behind ChatGPT), Erik Gafni, and Sasha Sheng. TypeSafe AI launched in September 2026 with a $40M seed round from DCVC. The model name comes from the 19th-century economist William Stanley Jevons.

Jev's architecture is not publicly documented. TypeSafe hasn't disclosed parameter count, base model, or training methodology. It handles text only, not images or video. The context window is 32K tokens on the hosted API, which is half of Clef's 64K.

Pricing: $0.042 per million input tokens. Output tokens are free.

Clef by Cloudflare

Clef is Cloudflare's first internally trained model, released as part of its Birthday Week announcements. There are two variants:

Clef is 27B parameters, post-trained from Qwen3.8-27B with its vision encoder intact. Multimodal: accepts up to four images per request alongside text. Context window: 64K tokens on Workers AI (the weights support 256K). Median latency on Workers AI: 209.3ms. Cloudflare benchmarks report BFCL case-exact accuracy of 98.5%.

Clef-Flash is 9B parameters, post-trained from Qwen3.5-9B. Also multimodal. Median latency on Workers AI: 38.8ms. Significantly faster than Clef at the cost of precision.

The weights for both are on Hugging Face under Apache 2.0. One clarification: Cloudflare described them as "open-source," but training data and pipeline are not published, so "open weights" is more accurate. You can download, self-host, and modify the weights; you cannot reproduce the training.

Pricing on Workers AI: Clef costs $0.24 per million input tokens; Clef-Flash costs $0.09 per million. A free daily allowance (10,000 neurons, roughly 700 average-comment analyses) applies when used via the Workers AI binding.

The test: 200 YouTube comments, five questions

200 random comments from the Better Stack YouTube channel were put through both models with identical prompts and five questions:

• Spam? (noul)
• Worth replying to? (noul)
• Says the video is wrong? (noul)
• Tone? (choice: praise, critique, joke, question, neutral)
• Toxicity? (score: 0-1)

Since there's no ground-truth label set for these comments, Claude Opus 5.5 served as the referee, answering the same five questions for each comment. Accuracy was measured as how often Jev or Clef matched the top referee answer with at least 50% confidence.

Speed

Jev was substantially faster. On average it returned results nearly twice as quickly as Clef.

Price

The 200-comment test cost significantly less on Jev than on Clef. Based on average comment length, Jev is 6.5x cheaper per comment than the standard Clef model.

The practical caveat: the entire Clef test run fell within Cloudflare's free daily allowance, making the effective cost $0. For small-scale use, the free tier changes the economics significantly.

Accuracy

Overall, Jev matched the referee 86% of the time and Clef matched it 83% of the time. A 3-point gap on 200 comments is narrow enough to treat as roughly equivalent overall.

The per-question breakdown tells a more interesting story:

Clef was more accurate on three of the five questions. Jev's overall lead came almost entirely from the Tone category, where it scored 75% versus Clef's 49%.

The cause of Clef's Tone weakness is identifiable: it defaults heavily toward "neutral." A comment like "Laya was released a year ago bro. Jev cloned on Laya" is a clear critique. Jev classified it as "critique" with 95% confidence. Clef classified it as "neutral" with 93% confidence. This tendency to collapse ambiguous sentiment to neutral flattens Clef's usefulness for anything requiring nuanced tone detection.

Test 2: vision analysis on 225 YouTube thumbnails

Clef's ability to analyze images opens up use cases that Jev can't reach. 225 YouTube thumbnails were run through Clef with two types of questions: predictive (will this perform above average?) and descriptive (what are the traits of this thumbnail?).

Predictive performance: a coin flip

Clef predicted 176 of 225 thumbnails would beat channel average. In reality, 49% of its picks actually outperformed. Its highest-confidence picks performed 1.3 times worse than typical videos.

Direct performance prediction from a static thumbnail is not something Clef can do reliably. This is an inherently difficult task and arguably the wrong question to ask any model.

Descriptive performance: genuine insight

Asking Clef to classify thumbnail traits rather than predict performance produced actionable data. By asking "what is the main visual?" and mapping classifications to actual view counts, a clear pattern emerged: thumbnails featuring a screenshot as the main visual received 1.6x more views than those featuring a diagram.

This is genuinely useful. The model can't tell you which thumbnail will win, but at scale it can help you discover which trait categories tend to correlate with performance. That's a meaningful use of multimodal classification that Jev can't access.

Which model fits which use case

Use Jev for: high-volume text classification where you need the lowest per-call cost and fastest latency. Spam detection, sentiment analysis, content routing, and any pipeline that processes thousands of items per hour. The 6.5x cost advantage compounds quickly at scale.

Use Clef for: anything involving images (thumbnail analysis, receipt classification, product photo tagging, screenshot understanding). Self-hosting requirements where you can't route data to TypeSafe's US-hosted API. Cloudflare Workers integrations where the free tier covers your volume. Cases where you need nuanced toxicity detection (Clef won on that category by 7 points).

Use Clef-Flash for: latency-critical hot paths where 38.8ms median matters more than maximum accuracy, and where you're already on Cloudflare's infrastructure. At $0.09/M tokens it splits the cost difference between Jev and full Clef.

Both models are genuinely new tools. Designing pipelines around structured probabilistic decisions rather than parsing LLM prose output is a real improvement for any workflow that runs decisions at volume.