Cloudflare AI Gateway pricing: Unified Billing, fees, limits

This page is built from Cloudflare's AI Gateway docs, model catalog, blog posts and status history, two public latency benchmarks, two write-ups by teams that use or studied it, a competitor's comparison page and user reports on the Cloudflare Community forum, all checked on October 21, 2026. Cloudflare's claims are labelled as its own.

What Cloudflare AI Gateway sells

Cloudflare AI Gateway started in September 2023 as a proxy in front of AI providers: you keep your own provider keys, send requests through a gateway URL, and get logs, analytics, response caching, rate limits, spend limits, retries and model fallbacks. These features are free on every Cloudflare plan. It supports 24 providers this way, from OpenAI, Anthropic and Google to Groq, Mistral, ElevenLabs, Replicate and OpenRouter.

The relay part is newer. Unified Billing, launched in closed beta on August 27, 2025, lets you buy Cloudflare credits and call models without your own keys. Over the HTTP API it covers OpenAI, Anthropic, Google AI Studio, Google Vertex AI, xAI and Groq; through the Workers AI binding it reaches any model in Cloudflare's catalog, which lists 262 models including Veo, Seedance and Flux. On August 7, 2026 Cloudflare merged Workers AI and AI Gateway into one binding, one /ai/ endpoint and one credit wallet.

Prices against the official APIs

Cloudflare's Unified Billing docs say inference is passed through with no markup, and that "a 5% fee is applied to all credits purchased". So 100 USD of credits costs 105 USD, and every model ends up about 5% dearer than at its maker:

Catalog prices come from Cloudflare's pages for Claude Sonnet 5.5, GPT-6 Sol and Gemini 4 Argon; the middle column is arithmetic from the 5% fee. With your own keys you pay the provider and Cloudflare charges nothing for routing.

Two billing details matter. The balance can go below zero "in rare instances", and Cloudflare then charges the card on file at the start of the next month. The docs state no minimum purchase, credit expiry or refund policy.

Limits that matter for a relay

• Unified Billing requests are capped at 200 per 60 seconds per gateway when Cloudflare's credentials are used; above that the gateway returns 429. The cap does not apply to your own keys. Free accounts can create 10 gateways and paid accounts 20.
• Logging of prompts and responses is on by default. Accounts with older gateways keep 100,000 logs in total on the free plan and 10 million per gateway on paid plans; accounts whose first gateway was created on or after September 24, 2026 follow Workers Logs limits.
• A "Require provider credentials" setting stops third-party requests from falling through to paid credits.

What public tests show

• ComputeSDK (a sandbox SDK company that does not sell a gateway) published an open benchmark on October 16, 2026, calling Cloudflare's passthrough route with a provider key, not Unified Billing. Over 100 requests each from Northern Virginia, median time to first token was 544 ms through Cloudflare against 394 ms direct on Claude Haiku 4.5 (Vercel: 462 ms), and 520 ms against 412 ms on gpt-5.4-mini. Every request succeeded. In that run Cloudflare added more delay than Vercel or OpenRouter.
• Edgee, a competing gateway, ran a competitor's benchmark from 43 Google Cloud regions in August 2026. Cloudflare had the lowest overhead of six gateways on Claude (26 ms on claude-sonnet-4-6) and the second lowest on gpt-5.4 (37 ms). The two tests disagree, which suggests the added delay depends on where you call from.
• Kalvium Labs, an AI engineering firm, wrote in April 2026 from client production systems that the gateway adds 20 to 60 ms depending on the distance to the nearest Cloudflare location, and that conversational cache hit rates are usually under 5%. It gives no method or sample.
• NxCode noted in an August 2026 desk analysis that paying with gateway credits raises Workers AI limits to 50 requests a minute for Kimi K2.6, Kimi K2.7 Code and GLM-5.2, from 20.

What users report

Cloudflare Community threads describe Unified Billing failures. One user says every third-party model request was rejected with error 2021 "Invalid User Credentials" while Cloudflare-hosted models worked. Another, with 10 USD prepaid and no own keys, got HTTP 402 on openai/gpt-image-2 while google/nano-banana-2 worked on the same gateway; a later thread reports 402 and 503 errors on OpenAI. One user says they paid over 1,000 USD and the balance did not move for more than seven days while an urgent case went unanswered. Others report top-ups failing with every card and PayPal, and a "Submit a case" link that led to an error page on the free plan. The forum blocks automated reading, so these threads' dates and any staff replies could not be checked.

MCP, Agent Skills and coding agents

Cloudflare runs several official MCP servers, including the Cloudflare API server (mcp.cloudflare.com/mcp), which covers its whole API and searches the docs. The separate AI Gateway MCP server, which searched gateway logs and showed prompts and responses, is listed as deprecated in Cloudflare's GitHub repo, with users told to move to the Cloudflare API server. The gateway's REST API docs do not mention passing MCP servers to models.

Cloudflare publishes official Agent Skills in its cloudflare/skills repo. The main cloudflare skill has an AI Gateway reference on configuration, SDK integration, billing and dynamic routing, and the repo installs as a plugin for Claude Code, Codex, Cursor and VS Code together with the Cloudflare MCP server. Anthropic's Skills API through the gateway is not documented.

There are setup guides for ten coding tools: Claude Code, Claude Desktop, Cursor, Gemini CLI, GitHub Copilot CLI, OpenAI Codex, OpenCode, Pi, VS Code and Xcode. Cline, Roo Code and Kilo Code have none. The REST API offers /ai/run for every model and modality, plus /ai/v1/chat/completions, /ai/v1/responses and /ai/v1/messages in OpenAI and Anthropic formats; provider-native URLs pass requests through unchanged. Cloudflare's own skill says the older unified /compat API is deprecated for single-model calls but still needed for dynamic routes. Cloudflare maintains ai-gateway-provider and workers-ai-provider for the Vercel AI SDK, and its docs publish an llms.txt.

Privacy, terms and who runs it

Cloudflare, Inc. is listed on the New York Stock Exchange and based in San Francisco. Logging can be turned off per gateway. Zero data retention routing exists only for Unified Billing requests on Cloudflare's credentials, and Cloudflare says it does not change gateway logging. Keys can be kept in Secrets Store, and DLP scanning and Llama Guard guardrails are available. Cloudflare publishes no AI Gateway uptime figure. On November 18, 2025 a network-wide Cloudflare outage caused widespread 5xx errors from 11:20 UTC until most traffic recovered around 14:30; the sources found do not say how AI Gateway itself was affected.

How it compares with Vercel, OpenRouter, Requesty and Eden AI

• Vercel AI Gateway charges no fee on credits and has no Vercel rate limit on paid credits, but even BYOK needs a positive credit balance there.
• OpenRouter adds 5.5% to credit purchases on its Standard plan and lists far more open-weight hosts.
• Requesty adds 5% to each top-up, the same as Cloudflare.
• Eden AI adds 5.5% and sells OCR, speech and translation APIs as well.

Who should skip it

Skip Unified Billing if you need more than 200 requests a minute per gateway on Cloudflare's credits, if a billing failure without a support channel you can reach would hurt, or if you want the lowest price: the 5% fee makes every model dearer than at its maker or at a no-fee relay. As a free logging, caching and limits layer in front of your own keys, especially for teams already on Workers, it is easy to justify.