Cloudflare Agents Pricing: What It Costs to Run an AI Agent

Cloudflare launched its unified Agents experience on August 4, 2026, giving developers one place to inspect hosted agents, replay sessions, and follow agent-aware traces. The important pricing detail is easy to miss: there is no single flat “Cloudflare Agents” fee. Your bill is assembled from the Cloudflare services each agent uses—Workers, Durable Objects, model inference, observability, and optional tools such as Browser Run or Sandbox.

This guide breaks down those layers so you can estimate a realistic monthly budget before moving an agent into production.

Cloudflare Agents pricing at a glance

As of August 5, 2026, these are the main cost layers documented by Cloudflare:

The Free plan can be enough for a prototype, but it behaves as a hard ceiling: when a daily free limit is reached, that operation fails rather than silently becoming a paid overage. A production service should normally use the $5 Workers Paid plan and enforce its own budgets.

What the Agents layer actually provides

Cloudflare Agents is a durable runtime rather than a new foundation model. Each agent session can have a persistent identity, local SQL state, real-time connections, scheduled work, and recoverable execution. A model still supplies the reasoning, while Workers and Durable Objects supply the application runtime.

Cloudflare’s current limits allow tens of millions of concurrent agent instances per account, up to roughly 250,000 definitions, and up to 1 GB of state for each unique agent. CPU time is limited to 30 seconds per agent event, but wall-clock time can continue while an agent waits for a database or model response.

That architecture matters for cost estimation. A long model response does not automatically mean a large Workers CPU charge, but keeping a Durable Object active when it could hibernate can create duration charges. Model tokens and external tools are billed separately from the agent runtime.

Cost layer 1: Workers requests and CPU

Every hosted agent starts with Cloudflare Workers pricing. The Paid plan has a $5 monthly minimum and includes 10 million Worker requests plus 30 million CPU milliseconds per month. Additional usage costs $0.30 per million requests and $0.02 per million CPU milliseconds.

Two details keep the runtime bill predictable:

• Cloudflare bills inbound Worker requests, not every subrequest the Worker makes to a model or API.
• Standard Workers pricing does not add separate data-transfer or bandwidth charges.

For many conversational agents, Workers itself will not be the expensive layer. Inference, browser time, paid tools, or a non-hibernating stateful session can cost more. Still, set a per-invocation CPU limit to reduce the risk of a retry loop becoming a denial-of-wallet incident.

Cost layer 2: Durable identity, state, and sessions

Each agent maps to a Durable Object, which is why a session can retain state and recover without rebuilding its context from an external database. Durable Objects pricing separates compute from storage.

On Workers Paid, the monthly allocation includes 1 million Durable Object requests and 400,000 GB-seconds of duration. Overage is $0.15 per million requests and $12.50 per million GB-seconds. New projects should use the recommended SQLite-backed storage, whose Paid allocation includes 25 billion rows read, 50 million rows written, and 5 GB-month of stored data.

The main cost trap is idle time. An object eligible for hibernation is not billed for duration while idle. A WebSocket or outbound connection configured in a way that prevents hibernation can keep the object billable. Cloudflare’s own pricing examples show a large difference between continuously active WebSockets and connections using the WebSocket Hibernation API.

For a chat or support agent, verify hibernation before optimizing tiny storage operations. It is usually the higher-impact control.

Cost layer 3: model inference

The starter Agents project uses Workers AI by default, although Cloudflare says developers can swap in OpenAI, Anthropic, Google Gemini, or another provider. If you bring an outside model provider, that provider’s token bill remains separate from Cloudflare.

Workers AI pricing includes 10,000 Neurons per day on both Free and Paid plans. Paid accounts are charged $0.011 per 1,000 Neurons beyond that daily allocation. Cloudflare also publishes equivalent token rates for each model. For example, its August 5 table lists Qwen3 30B A3B FP8 at $0.051 per million input tokens and $0.335 per million output tokens, while Gemma 4 26B A4B is listed at $0.10 per million input tokens and $0.30 per million output tokens.

Do not estimate inference from request count alone. Record these four values:

1. Agent turns per month.
2. Average input tokens per turn, including retrieved context and history.
3. Average output tokens per turn.
4. Extra model calls for planning, evaluation, reranking, or subagents.

A “single” user request may trigger several model calls. Session replay and traces are useful because they reveal retries, subagent calls, and oversized context that a request counter cannot show.

Optional costs: browser, sandbox, and observability

Agents that only call APIs are usually cheaper than agents that operate websites or execute arbitrary software.

Browser Run

Browser Run pricing gives Paid accounts 10 browser hours per month, then charges $0.09 per additional hour. Direct browser sessions also include 10 average concurrent browsers, with each additional average concurrent browser costing $2 per month. Cloudflare calculates concurrency from the monthly average of each day’s peak, so brief daily peaks still affect the average.

Prefer direct APIs when available. Use a browser only when the target workflow genuinely requires visual navigation or a site has no suitable API.

Sandbox

Cloudflare’s Sandbox SDK is billed through the underlying Containers platform, with Workers and Durable Objects charged as related services. This makes a code-execution agent a multi-meter workload. Apply timeouts, restrict network access, cap filesystem growth, and require approval for sensitive actions. NextPJ’s AI agent sandbox security checklist covers the operational controls that belong beside the budget controls.

Agent traces

The August 4 Cloudflare Agents announcement says agent tracing is free during beta. Starting October 1, 2026, each span will count as a Workers Observability event. The announced allowances are 200,000 events per day with three-day retention on Free, and 20 million events per month with seven-day retention on Paid; Paid overage will be $0.60 per million events.

Tracing can record model calls, tool execution, approval events, token usage, and supported subagent activity. That visibility is valuable, but message and tool payloads may contain personal data or secrets. Disable payload capture where needed, sample routine traffic, and export only what your retention policy permits. For implementation context, see NextPJ’s guide to tracing AI agents with OpenTelemetry.

A practical monthly estimation method

Build the budget in this order:

1. Establish the floor. Use $0 for a capped prototype or the $5 Workers Paid minimum for production.
2. Estimate inference. Multiply monthly input and output tokens by the selected model’s current rates, then account for the daily Workers AI allocation if applicable.
3. Measure state activity. Count Durable Object requests, duration, reads, writes, and stored data. Confirm that idle sessions hibernate.
4. Add expensive tools. Track browser hours, average browser concurrency, container runtime, paid MCP tools, and third-party API charges independently.
5. Budget telemetry. Estimate spans per turn before tracing becomes billable on October 1, 2026.
6. Add a safety margin. Reserve 20% to 30% for retries, traffic bursts, evaluations, and model-routing changes.

Do not use one blended “cost per request” until you have production traces. First separate model, runtime, state, tools, and telemetry; then divide the total by successful completed tasks. Cost per successful task is more useful than cost per HTTP request.

Production checklist before you scale

• Set account alerts and application-level daily spend limits.
• Cap steps, retries, output length, browser time, and container lifetime.
• Require human approval before purchases, publishing, deletion, or permission changes.
• Test Durable Object hibernation with realistic WebSocket traffic.
• Sample traces and redact sensitive payloads before export.
• Compare at least two suitable models on task success and total cost, not token price alone.
• Add a kill switch that disables tool use without taking the whole service offline.
• Review pricing pages again before launch; beta terms and model rates can change.

Cloudflare also announced programmable Wallets on August 4, but the Wallets announcement uses future-tense language for funding, payments, virtual wallets, and spending guardrails. As of August 5, users can claim a wallet handle; teams should not treat autonomous API payments as a generally available production control yet.

Is Cloudflare Agents cost-effective?

Cloudflare Agents is a strong fit when your application benefits from durable per-user state, globally distributed sessions, WebSockets, scheduling, or tight integration with Workers. Its billing is especially attractive when sessions hibernate correctly and the agent relies on APIs rather than long browser or container jobs.

It is less simple when a workflow needs heavy sandbox compute, long browser sessions, or an external premium model on every step. In those cases, Cloudflare may still be a good control plane, but the $5 runtime floor tells you little about the final bill.

The safest approach is to deploy a narrow task, trace it, and calculate cost per successful completion. Cloudflare’s new Agents dashboard makes that measurement easier; disciplined limits are what make it safe to scale.