Cloudflare launches Clef as TypeSafe's 'decision model' idea spreads across big tech

What happened? On 1 October 2026 Cloudflare released Clef and Clef-flash, its first models trained in-house by the Workers AI team. They are "decision models": instead of generating text, they take a state (a support ticket, a web page, an agent trace) plus a set of typed questions and return calibrated probabilities over predefined answers. Clef is post-trained on Qwen3.8-27B and Clef-flash on Qwen3.5-9B; both add a vision encoder and a 64k context window, are fully API-compatible with Jev, and are open-sourced on Hugging Face under Apache 2.0. Cloudflare reports Clef-flash at a median latency of about 39ms, and says Clef leads the Jev Decision Index. It also launched a reinforcement learning fine-tuning service, delivered first through its forward-deployed engineers and later as a self-serve platform built on AI Gateway, Containers and the Replicate bring-your-own-model stack. OpenClaw creator Peter SteinbergerDealroom has a profile for this one. Try Dealroom → shared the launch with the comment: "Never seen an idea spreading so fast."

The idea: the concept comes from TypeSafe AIDealroom has a profile for this one. Try Dealroom →, founded by former OpenAI researcher Diogo Almeida, who helped develop the RLHF work behind ChatGPT. On 15 September 2026 it launched Jev, a "System One model" named after economist William Stanley Jevons. Jev gives up string generation entirely, so its outputs are type-safe by construction and cannot hallucinate, and it is trained with a method TypeSafe calls Reinforcement Learning for Calibrated Decisions (RLCD). TypeSafe prices input at $0.042 per million tokens with free output and quotes 70-500ms response times, 40-200x faster than frontier LLMs on decision-shaped tasks. The pitch is "smart if-statements" for software: routing, classification, scoring, guardrails and agent monitoring. Demand briefly overwhelmed TypeSafe's API, and Vercel reported 5-18x faster and more accurate results after swapping an OpenAI model for Jev in a command-safety classifier.

• OpenAI (29 September): OpenAI previewed a Decisions API powered by GPT-6 Luna for classifying content, routing requests and choosing an agent's next action.
• Liquid AI (29 September): Liquid AI announced d1, its first decision model, claiming to be the first to beat Jev on Hugging Face's Decision Index.
• Amazon Web Services (1 October): AWSDealroom has a profile for this one. Try Dealroom → open-sourced Strands Decider 2B, built on Qwen3.5-2B. It began as a side project by distinguished engineer Marc Brooker after he saw Jev.
• Perplexity (1 October): Perplexity launched a Decisions API on its open-weight pplx-decider-v1-27b, priced at $0.04 per million input tokens.
• Open source: TechCrunch counts dozens of community variants, including Kev 9B and Laya, alongside Clef.

Why it matters: decision models point to a cheaper layer of AI for automation and agents, sitting between hand-written rules and full LLMs. The speed of copying also raises the question of a moat. Almeida told TechCrunch he sees no real competition yet, arguing the current batch is "ML people wanting to implement a cool architecture". Early developer comparisons after Cloudflare's launch were mixed: several found Clef more expensive per token than Jev and not always more accurate on their own tasks, despite Cloudflare's benchmark lead.

Read more: Cloudflare · Peter Steinberger on X · TypeSafe AI · TechCrunch · TechCrunch · OpenAI Devs · Perplexity Devs · Liquid AI