Cloudflare Workers AI | Ry Walker Research

Key takeaways

Workers AI connects Cloudflare applications and external HTTP clients to managed GPU inference.
Billing and access depend on the model; premium models have separate paid-plan or prepaid-credit requirements.
LoRA deployment remains a selective beta, and API compatibility needs testing against the application's actual calls.

FAQ

What is Cloudflare Workers AI?

A managed inference service accessible through Workers and Pages bindings or HTTP APIs, with a catalog spanning text, embeddings, images, and audio.

Is Workers AI free?

Eligible usage includes 10,000 neurons daily. Paid overage is $0.011 per 1,000 neurons, while designated premium models require paid access.

Can Workers AI run custom models?

Selected models accept externally trained LoRA adapters through a beta. Arbitrary model requirements require a separate conversation with Cloudflare.

Does OpenAI compatibility guarantee coding-agent compatibility?

No. Cloudflare documents compatible endpoints, including a Responses example, but applications still need model-specific streaming and tool-call tests.

Executive Summary

Cloudflare Workers AI is a generally available inference service that runs a curated model catalog on managed GPUs. Applications can call it from Workers or Pages, or use HTTP APIs from other infrastructure. It belongs in the AI Inference Platforms comparison as an inference provider, distinct from the company's sandbox execution service.^[1]

Its strongest fit is an application that already uses Cloudflare and wants inference without provisioning a GPU fleet. The important qualification is that availability, billing, and limits vary by model. Workers AI should be evaluated with the intended model and client protocol, rather than assumed to be an unrestricted host for any weights or an interchangeable replacement for every model API.^[2]^[3]

Attribute	Verified position
Operator	Cloudflare^[1]
Service status	Core service generally available; LoRA adapters remain beta^[1]^[4]
Deployment	Managed inference through bindings and HTTP APIs^[1]
Review date	September 15, 2026; documentation review, not a hands-on benchmark

Product Overview

Key Capabilities

Capability	Practical scope
Inference catalog	Text generation, embeddings, image, and audio models; inspect each model's task and input schema^[1]
Application binding	Call `env.AI.run` from a Worker with a configured AI binding; Pages Functions can also bind the service^[5]
Compatible APIs	Documented chat-completion and embedding endpoints; the guide also includes a GPT-OSS Responses example^[3]
LoRA adapters	Upload adapters for selected supported, non-quantized base models; this is adapter deployment rather than an in-product training workflow^[4]

Product Surfaces

Surface	Description	Availability
Workers binding	Inference within a Cloudflare application	Documented self-service workflow^[5]
HTTP API	Inference from an external client	Account credentials and model access required^[3]
Custom requirements	Higher limits or private/custom-model needs	Contact Cloudflare; do not infer self-service arbitrary-weight deployment^[2]

Technical Architecture

The application runs separately from the model-serving GPU infrastructure. A binding or authenticated API request selects the model and supplies inputs; Cloudflare operates the inference service. This removes the customer's GPU provisioning work while retaining dependencies on the provider's catalog and per-model capacity policy.^[1]

For a Worker, add an AI binding to the existing Wrangler configuration:

{
  "ai": { "binding": "AI" }
}

An inference call inside its request handler can then look like this:

const result = await env.AI.run("@cf/google/gemma-4-26b-a4b-it", {
  messages: [{ role: "user", content: "Explain a database index in one sentence." }],
  chat_template_kwargs: { enable_thinking: false }
});
return Response.json(result);

This is an adapted documentation example, reviewed but not executed for this report. Wrangler development still calls the account's remote inference service and consumes billable usage; a local application URL does not imply local model execution.^[5]

Key Technical Details

Detail	Consequence
Model-specific API behavior	Test structured output, streaming events, tool calls, and cancellation with the actual client^[3]
Rate limits	Task defaults have model exceptions; some premium models have different standard and prepaid-gateway limits^[2]
Adapter lifecycle	LoRA uploads require configuration and weights files; documented limits include 300 MB per adapter and 100 adapters per account^[4]
Data policy	Cloudflare says customer content is not used to train or improve models without consent; application-selected storage remains a separate decision^[6]

A no-training commitment is not itself a claim about universal zero retention, a selected geographic region, or every connected storage product. Teams with those requirements should establish the applicable processing and storage terms before deployment.^[6]

Strengths

Short integration path for existing Workers applications. The binding keeps inference calls inside the application's existing deployment workflow.^[5]
External access is supported. The compatible HTTP surface also serves applications hosted elsewhere.^[3]
A documented customization path exists. Supported LoRA adapters can specialize selected models without the customer operating a GPU server.^[4]

Cautions

A catalog is an architectural constraint. Confirm the desired model and any custom-weight requirements before choosing the platform.^[2]
LoRA remains selective and beta. Adapter support does not establish support for every base model, architecture, or training method.^[4]
Account limits affect concurrency. Capacity planning must use the model's limits, rather than a single platform-wide requests-per-minute assumption.^[2]
Model terms remain relevant. Cloudflare's data policy assigns responsibility for the selected model's licensing terms to the customer.^[6]

What Developers Say

On April 12, 2026, Hacker News user mrnoname, the author of a Codex-to-Workers-AI proxy, described building it to spend existing Cloudflare startup credits. They reported that the Responses interface did not map cleanly to their Codex workflow and that Kimi K2.5 had worked for roughly a day through the proxy. This is a useful first-hand integration report with a stated incentive, not a reliability study.^[7]

The current official compatibility guide includes a Responses example. The April report therefore supports testing the full client interaction; it does not establish that the same incompatibility remains in September.^[3]

Pricing & Licensing

Published terms checked September 15, 2026:

Usage	Published terms
Eligible free allowance	10,000 neurons per day
Paid neuron usage	$0.011 per 1,000 neurons above the allowance
GPT-OSS-120B example	$0.35 per million input tokens; $0.75 per million output tokens
Premium models	Designated models require Workers Paid or prepaid AI Gateway credits

The pricing page maps model-specific units to the billing system. Do not treat a neuron as a token or apply one model's conversion to another.^[8] LoRA usage is described as free during its beta, which is a temporary feature term rather than a durable price guarantee.^[4]

Licensing model: Hosted service with model-specific licensing obligations.^[6]

Hidden costs: Account for application hosting, chosen storage, repeated development requests, and the model's actual output volume. The inference allowance does not imply that the rest of the application stack is free.^[5]^[8]

Competitive Positioning

Direct and Adjacent Alternatives

Alternative	Decision-relevant distinction
Replicate	Public model APIs and private Cog deployments offer another customization path. Private-model setup and idle time can be billed.^[9]
Modal	Shared token-priced endpoints coexist with dedicated endpoints and custom weights billed for compute.^[10]
Tembo	A coding-agent platform whose Max gateway also exposes model access to external tools; relevant when inference is part of an agent workflow rather than only an application endpoint.^[11]^[12]

Disclosure: Ry Walker is Tembo's founder and CEO. No native Tembo–Workers AI integration was verified in this review.

Replicate announced that it had joined Cloudflare on December 1, 2025. Common ownership does not establish interchangeable catalogs, APIs, or billing between Replicate and Workers AI.^[13]

Choose Workers AI when its catalog and account limits fit an application already built around Cloudflare bindings. Evaluate a dedicated deployment alternative when custom weights, capacity ownership, or a serving configuration absent from the catalog is a requirement.

Ideal Customer Profile

Best fit: Teams building Cloudflare applications, HTTP clients using supported models, and developers evaluating a documented LoRA path.

Poor fit: Workloads that require an unverified model, rely on guaranteed capacity without an agreement, or assume that an API-compatibility label covers every coding-agent protocol detail.

Viability Assessment

Current documentation exposes a usable core service, model-specific pricing, limits, and maintained integration paths. Cloudflare's completed Replicate acquisition also makes the surrounding product portfolio broader, but it does not prove a particular Workers AI adoption level or future catalog commitment.^[1]^[13]

The practical long-term question is portability: retain representative inference requests and tests so a model change, pricing change, or provider migration can be evaluated against the application's actual requirements.

Bottom Line

Workers AI is a credible managed inference option for applications that benefit from Cloudflare's deployment environment and fit its model catalog. Its value is the integration and operating model; a universal speed or cost ranking would require workload-specific measurements this review did not perform.

Recommended for: Cloudflare application teams with supported models and measured request patterns.

Not recommended for: Buyers assuming arbitrary model hosting or untested end-to-end protocol compatibility.

Outlook: Evaluate catalog, adapter status, and capacity terms at deployment time; the service and its beta features have different maturity levels.

Research by Ry Walker Research • methodology

Sources