Most AI projects die at the infrastructure stage: you need an API key, a credit card, or a GPU server before you can ship anything. Cloudflare Workers AI quietly removes that barrier. On a free Cloudflare account you get access to dozens of open-weight models — Llama 3.3 70B, Llama 3.1 8B, Llama 4 Scout, Qwen Coder 32B, Whisper, FLUX-1-schnell and more — all callable from serverless code running on Cloudflare's edge network. Here are the eight best things you can actually build on that free tier in 2026.
How the free tier works
Before the use cases, the numbers. Cloudflare prices Workers AI in "neurons", a unit that normalizes compute across model sizes. The free tier gives you 10,000 neurons per day, resetting daily, as an account-level budget. Text generation is also subject to roughly 300 requests per minute. In practical terms that is on the order of 10,000 short Llama 3.1 8B completions per day, or 100,000+ embedding generations — far more than most prototypes, side projects, or MVPs ever need.
There are two practical catches. First, inference happens on Cloudflare's network, not your hardware, so it is not quite as instant as a local model — simple chat answers typically land in a few to ten seconds. Second, a small number of models require a billing method on file even though they still draw from the free quota, so check each model's page in the Cloudflare catalog before you build around it. With that understood, here is what the free tier is genuinely good at.
1. Chatbot endpoints without a backend bill
This is the headline use case. A Cloudflare Worker can expose a chat endpoint powered by Llama 3.3 70B or the faster Llama 3.1 8B, with no separate server and no API key management. The model lives behind an AI binding in your wrangler.toml:
[ai]
binding = "AI"Then a few lines of TypeScript call env.AI.run("@cf/meta/llama-3.3-70b-instruct-fp8-fast", { messages }). Indie developers are already replacing OpenAI calls in prototypes this way: support bots for small SaaS products, FAQ assistants for documentation sites, and internal Q&A tools. If you gate the endpoint — a per-IP daily cap stored in Workers KV is the standard recipe — a free account can serve hundreds of real conversations a day at zero cost.
2. Semantic search for your own content
Full-text search returns keyword matches; semantic search returns meaning matches. With Workers AI you get both halves of a semantic search stack for free: the bge-m3 (or embeddinggemma) model generates embeddings, and Cloudflare's Vectorize service stores and queries the index. This combination is the path of least resistance for adding "search that understands" to a docs site, a blog, or a product catalog — you embed your pages once, then every query becomes a nearest-neighbor lookup at the edge. It is one of the cheapest per-request workloads in the whole catalog, which is why the free neuron budget stretches so far here.
3. Content moderation before content hits your database
Moderation is a classification problem, and the 8B-class models on Workers AI are good at binary and category judgments. Two options exist: general models like Llama 3.1 8B prompted to classify, or the dedicated llama-guard-3-8b moderation model. The edge architecture is the real win — a Worker intercepts the submission, runs the classification, and only writes to your database if the content passes. That keeps toxic posts, spam, and prompt-injection attempts out of your storage and your UI without adding a moderation vendor to your stack.
4. Image generation on the free tier
FLUX-1-schnell is available through Workers AI, which means free-tier users get AI image generation with no GPU of their own. Neuron pricing is higher per image than per chat message (costs are quoted per tile plus per step), so this is not a service for infinite image feeds — but it is perfect for user avatars, generated thumbnails, OG images for blog posts, or placeholder art in an MVP. Several indie projects generate their marketing images this way and never pay for a separate image API.
5. Speech-to-text with Whisper
Whisper's neuron cost is quoted per audio-minute, and the free daily budget covers plenty of short clips. Practical uses: transcribing voice notes in a support app, generating captions for short videos, or voice-input features in a mobile web app. whisper-tiny-en is there for fast cheap English transcription, while whisper-large-v3-turbo handles accuracy-critical jobs. Combined with a Worker that accepts uploads, you get an audio transcription endpoint with no server and no audio pipeline to maintain.
6. Translation and summarization middleware
Translation models like m2m100-1.2b and summarization via the Llama models make Workers AI an excellent "middleware" layer: a Worker sits in front of your API or origin, rewrites responses, and sends them on. Real patterns in the wild include translating product descriptions into multiple languages at request time, summarizing long articles into preview snippets for feeds, and extracting structured data (prices, dates, names) from unstructured text before it enters a database. Because the Worker and the model are co-located on the edge network, the added latency of the AI call is minimized.
7. AI coding agents and dev tools
This one surprises people. The catalog includes Qwen Coder 32B and the Llama 4 Scout model, which testers report as genuinely useful for code completion, documentation generation, and code review tasks. Developers are wiring these models into side-project dev tools: GitHub PR reviewers that comment on diffs, commit-message generators, and internal coding assistants that run behind a team webhook. Since coding requests are short and intermittent, a single free account's daily budget covers a surprising amount of real dev-tool usage.
8. Voice interfaces and vision tasks
Round out the list with the smaller models. melotts gives you text-to-speech for audio versions of articles or voice prompts in a kiosk-style app. llava-1.5-7b (image-to-text) powers automatic alt-text generation, image captioning for CMS uploads, and "describe this screenshot" debugging tools. None of these are frontier models, but on the free tier they turn single-feature ideas into shippable products with zero model-hosting work.
Setting up your first use case
The mechanics are the same for all eight:
1. Install the Wrangler CLI: npm install -g wrangler
2. Log in: wrangler login
3. Add the AI binding to wrangler.toml (see the config snippet above)
4. Call the model in your Worker: await env.AI.run(modelId, input)
5. Deploy: wrangler deploy
The critical step is step zero — measure. Fire a handful of real calls, check how many neurons each costs in the Cloudflare dashboard, and confirm your expected traffic fits inside 10,000 neurons/day. Then add guardrails: per-IP rate limits, a global daily cap set comfortably below the free quota, and input length validation. A public endpoint with no caps will burn the daily budget by breakfast.
When the free tier is not enough
Honesty matters here. If your app needs thousands of long conversations per day, the largest 120B-class models at scale, or sub-second latency on every request, you will outgrow the free budget. Cloudflare's paid tier is $5/month for the Workers Paid plan, with AI priced at $0.011 per 1,000 neurons — modest, but real. Also note the catalog is not infinite: it is a curated set of open-weight models, so if you need a specific proprietary model, you will still reach for its own API. For everything else — prototypes, MVPs, internal tools, and modest-production features — the free tier is the most generous inference offer on the internet right now.