Cloudflare Workers always-free tier: 100,000 requests per day

Cloudflare’s Workers Free plan includes 100,000 requests per day with 10 ms of CPU time per invocation. On top of that, Workers AI includes “10,000 Neurons per day at no charge,” and that daily AI allowance is the same whether you are on the Free or Paid plan. Both allowances reset daily at 00:00 UTC. There is no dollar credit and no expiring trial; this is an always-free tier. The official pricing pages do not state a card requirement for the free plan. Checked on 2026-09-02.

What the free plan covers

The 100,000 daily requests cover Workers and Pages Functions. The free plan also includes 100,000 KV reads per day and 100,000 Hyperdrive queries per day, and requests for static assets are free outright. The 10 ms CPU limit is per invocation, which suits routing, API glue, and light transformation work rather than heavy compute.

Workers AI and the Neuron math

Neurons are Cloudflare’s cross-model usage unit. You get 10,000 per day free; beyond that you need Workers Paid, where overage is $0.011 per 1,000 Neurons and the plan minimum is $5 per month including 10 million requests. One gate worth knowing: the pricing page’s note lists five models that require a paid billing method, kimi-k2.6, kimi-k2.7-code, glm-5.2, deepseek-v4-flash-0731, and deepseek-v4-pro-0813. Those run with Workers Paid or prepaid AI Gateway credits, not on the free allowance alone. Workers AI pricing was updated 2026-08-18, so recheck before relying on the overage rate.

What happens at the limit

Exceed any daily limit and that operation type fails until the reset. Nothing is billed on the free plan; it just stops. That makes the free tier safe for a wallet and unsafe for anything that must not go down, the same trade the Render free tier makes with suspension at quota.

Choose or skip

Choose Workers Free if your workload is many small requests and you want a generous daily allowance that never expires. Skip it if a single invocation needs more than 10 ms of CPU or your models sit behind the paid gate.

What 10,000 Neurons a day converts to on named models

The pricing page publishes a Price in Neurons column per model, so the daily allowance can be converted into tokens with plain arithmetic, computed 2026-08-24 at the listed rates. On @cf/meta/llama-3.2-1b-instruct, listed at 2457 neurons per million input tokens and 18252 neurons per million output tokens, 10,000 Neurons is about 4.07 million tokens input-only, or about 548,000 tokens output-only. On the larger @cf/meta/llama-3.1-8b-instruct-fp8-fast, listed at 4119 neurons per million input and 34868 neurons per million output, the same allowance is about 2.43 million input tokens or about 287,000 output tokens. On the @cf/baai/bge-m3 embeddings model, listed at 1075 neurons per million input tokens, it is about 9.3 million tokens embedded per day. Real prompts mix input and output, so treat these single-sided numbers as ceilings from the official rate card, not as a promise of daily throughput.

What I have done with it

Not tested yet. This page reports only what the official pages state.

For neighboring free tiers, see Deno Deploy and Vercel, or read how to choose compute. If a single request needs more than the 10 ms of CPU this plan allows, the neighboring shape is a small always-on container rather than an edge function: Koyeb keeps one free web Service running, with a card required and a signup that defaults to its paid Pro plan until you downgrade. Every grant I track is on the credits index.