Cloudflare Clef: Decision Models for Enterprise AI Agents
On October 1, 2026, Cloudflare released Clef and Clef-flash, the first AI models trained entirely by its Workers AI team. On October 2, it followed with a Web Search API for AI Gateway, integrating live search from Ceramic.ai, Exa, and Linkup into the same control plane.
The two launches, taken together, are the clearest statement yet of Cloudflare’s agent cloud thesis: that the right place to run enterprise AI workloads is not inside a hyperscaler’s managed AI service, but distributed across the network edge, with decisions, memory, search, and browser automation all governed through a single plane.
Why Decision Models Are Eating Agent Routing
Decision models are not a new category, but the market is moving fast. TypeSafe AI’s Jev launched in mid-September 2026 and reached 13 percent of paid teams on Vercel’s AI Gateway within 24 hours. OpenAI shipped a Decisions API preview at DevDay on September 29. Now Cloudflare has its own entrant.
The core idea is simple. When an enterprise agent needs to route a support ticket, classify a domain for threat intelligence, or decide whether to escalate a transaction, calling a full frontier LLM like GPT-6 or Claude is expensive and slow. A decision model does the same job in milliseconds at a fraction of the cost.
Cloudflare’s own internal measurement puts numbers to that gap. Using Clef paired with Browser Run to fetch, render, and classify a live domain for threat intelligence: 2.2 seconds. Running the same workflow with gpt-oss-120b, its fastest general-purpose LLM: 4.7 seconds. Clef also returned more classification categories per request. That 2x latency advantage compounds across high-volume agent pipelines where hundreds of routing decisions run in parallel.
Clef and Clef-Flash: Architecture and Specs
Clef ships as two models with different performance profiles.
| Model | Parameters | Context Window | Optimal Use | Input Price |
|---|---|---|---|---|
@cf/cloudflare/clef | 27B | 64K tokens | High-precision decisions | $0.24 / M tokens |
@cf/cloudflare/clef-flash | 9B | 64K tokens | Low-latency hot-path decisions | Lower than Clef |
Each model accepts a state, which can be plain text, structured JSON, or up to four embedded images (PNG, JPEG, WebP), alongside a map of up to 64 typed questions. Responses return a probability for every allowed answer under each question ID. Three question types are supported:
- bool: returns a probability that the answer is yes
- choice: returns the selected option with per-option confidence scores
- score: rates content against an ordered rubric, returning a probability-weighted score and per-level probabilities
The inclusion of a vision encoder is what most distinguishes Clef from TypeSafe’s Jev, which handles text only. An enterprise compliance team can pass a document page alongside a classification rubric and get a structured decision in one call. A fraud-detection workflow can score a transaction with an attached receipt image. The vision capability does not require a separate multimodal pipeline.
Both models are available immediately on Workers AI and as open weights on Hugging Face under Apache 2.0. They implement the System One API introduced by TypeSafe, so any existing Jev integration can migrate by changing the endpoint URL and model identifier, with no schema rewrites required.
According to the Cloudflare launch post, Clef currently leads the Jev Decision Index. Independent third-party benchmark validation has not been published yet. Enterprise teams should run their own evaluations on representative workloads before committing.
Reinforcement Learning Fine-Tuning: Proprietary Classifiers at the Edge
The Clef launch bundles a new reinforcement learning fine-tuning service. The pipeline connects existing Cloudflare primitives into an end-to-end training loop:
- AI Gateway captures all decision traffic and automatically assembles a labelled dataset.
- Workers AI generates rollouts against the base Clef model.
- Cloudflare Containers run a sandboxed RL environment for scoring and replaying agent actions.
- A new Trainer component updates model weights.
- The fine-tuned model redeploys to Workers AI alongside the base version.
The service launches as a hands-on engagement with Cloudflare’s forward-deployed engineer team. Customers identify a workflow where they have years of labelled decisions, and the FDE team guides the data capture, training, and deployment cycle. A self-serve platform is on the roadmap.
This matters for enterprises that cannot compete with general-purpose benchmarks but do hold irreplaceable proprietary data. A large insurer classifying claims types, a marketplace flagging policy violations, a B2B SaaS triaging support tickets: each of these has domain-specific labelled history that a generic Clef model will never match without retraining. The RL loop makes that investment possible without standing up a separate ML platform.
Cloudflare notes it is already using this approach internally, training Clef variants for Bot detection, Trust and Safety review, and Support triage, all areas where fifteen years of network data creates a training advantage that no third party can replicate.
Web Search API: Grounding Agents in Live Information
On October 2, one day after the Clef release, Cloudflare announced Web Search API via AI Gateway with three launch partners: Ceramic.ai, Exa, and Linkup.
The problem the product addresses is familiar to anyone who has debugged a production agent. When a model needs live data, it typically guesses a URL and fetches it directly. If the URL is wrong or stale, the agent gets a 404 or outdated content and proceeds anyway. Web search gives agents the same entry point humans use: a query, followed by relevant structured results.
The integration sits inside AI Gateway, which means:
- Search queries draw from the same credit balance as model inference, unified billing across the full AI stack.
- Queries appear in AI Gateway observability logs alongside LLM calls, giving teams a complete trace of what information agents acted on.
- Access controls and BYOK (bring your own key) work the same way as for any other AI Gateway provider.
All three launch partners have committed to Cloudflare’s bot crawling standards: honest crawler identification, robots.txt compliance, and inclusion of source URLs in every result. For enterprise legal and compliance teams, this sourcing transparency reduces the risk that AI-generated outputs cannot be traced back to their factual basis.
Cloudflare plans to add web search as a native Server Tool inside AI Gateway, making it available to any agent pipeline as a built-in capability rather than a custom tool definition. That roadmap step would put search alongside Clef, browser automation from Kitesurf, and agent identity and payments as native primitives in the same platform.
What This Means for Enterprise AI and GTM Teams
Clef and the Web Search API are not standalone experiments. Read alongside Cloudflare OS (agent workspace framework), Kitesurf (browser-use for agents), and Wallets (agent payments), Cloudflare is building a vertically integrated layer-zero for agentic enterprise workloads.
The architectural bet differs from hyperscaler offerings in one important way: latency gravity. Cloudflare’s 300-plus network locations mean that edge-hosted agent decisions can run near the data sources and users they serve, avoiding the round-trip costs that accumulate when an agent in us-east-1 makes decisions on behalf of users in Singapore or Frankfurt.
For enterprise AI and GTM teams evaluating agent infrastructure, the October launches add two concrete capabilities to the Cloudflare stack:
- Fast, typed routing decisions for any classification task, with an open-source model they can run locally, a hosted version with enterprise data guarantees, and a fine-tuning path for proprietary domain adaptation.
- Live web context integrated into the same Gateway control plane as model inference, with sourced, standards-compliant crawling from three independent search providers.
Teams building on Workers AI can access both features today without a new vendor contract. For teams on other infrastructure, the open-source weights and Jev-API compatibility lower the migration threshold considerably.
The decision model category is moving fast. TypeSafe’s Jev raised $1 billion at a $10 billion valuation in late September, and OpenAI’s Decisions API is days from a broad release. Clef enters the race with the latency and cost advantages of edge hosting, Apache 2.0 licensing, multimodal support, and a fine-tuning path that turns proprietary decision history into a durable model advantage.