Workers AI Models and Pricing

Practical Workers AI model selection, binding usage, and cost calculation

Workers AI is the Cloudflare-hosted inference path that sits directly beside Worker code. Use it when the application can run on Cloudflare-hosted models and the call belongs in the Worker request flow. Move through AI Gateway when the app needs external providers, provider observability, or controls that span multiple providers.

Binding Basics

Expose Workers AI to a Worker with an AI binding:

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

Then call a model with env.AI.run(modelId, input). The input shape depends on the model, but text generation normally uses messages:

export interface Env {
  AI: Ai;
}

const modelId = "@cf/meta/llama-3.3-70b-instruct-fp8-fast";

export default {
  async fetch(_request, env): Promise<Response> {
    const input = {
      messages: [{ role: "user", content: "Summarize Workers AI in one sentence." }],
    };

    const result = await env.AI.run(modelId, input);
    return Response.json(result);
  },
} satisfies ExportedHandler<Env>;

For local development, use remote Workers execution for real model calls. There is no local GPU inference hidden inside wrangler dev.

Model Categories

Do not treat the model catalog as one long list. Pick by task first, then choose the smallest model that passes your quality bar.

• Text Generation: chat, summarization, extraction, code, tool use, and structured judgement.

• Text Embeddings: turn text into vectors for Vectorize, semantic search, recommendations, deduping, or RAG.

• Text-to-Image: generate or edit images from prompts and image references.

• Image-to-Text / vision: caption images, answer questions about screenshots or photos, and classify visual inputs.

• Automatic Speech Recognition: transcribe audio or prepare voice-agent input.

• Text-to-Speech: turn application text into generated speech.

• Text Classification / reranking: sentiment, safety, relevance scoring, or query/document reranking.

• Translation: translate text between supported language pairs.

Practical Text-Generation Picks

Checked on 2026-07-08 against the official Workers AI model pages and pricing page. This is a short starting set, not a ranking and not a catalog dump. Recheck the official docs before locking production cost assumptions.

The practical default is usually not the largest model. Start with a cheaper candidate for the exact task, measure failures, and escalate only the requests that need stronger reasoning, longer context, tool use, or vision.

Cost Calculation

Cloudflare bills Workers AI in Neurons. The pricing page also shows token prices as an equivalent view so you can compare models with normal LLM cost math.

Use this estimate for text-generation calls:

input_tokens / 1_000_000 * input_price + output_tokens / 1_000_000 * output_price

For example, if a model costs $0.051 per million input tokens and $0.335 per million output tokens, then a request with 20_000 input tokens and 2_000 output tokens is roughly:

20_000 / 1_000_000 * 0.051 + 2_000 / 1_000_000 * 0.335

That token view is for planning. The billable backend unit is still Neurons. Workers Free and Workers Paid include Workers AI, and the daily free allocation is 10,000 Neurons/day. On Workers Paid, usage above that allocation is charged at the current neuron rate; checked on 2026-07-08, that was $0.011 / 1,000 Neurons. Pricing and model availability drift, so treat copied prices as dated notes, not durable constants.

JSON Mode For Judgement Outputs

For classification, extraction, moderation, routing, or other judgement outputs that feed automation, prefer JSON Mode on a supported model. Add response_format with a JSON Schema, keep the schema narrow, and validate the returned object in your own code before acting on it.

const response = await env.AI.run("@cf/meta/llama-3.3-70b-instruct-fp8-fast", {
  messages: [
    {
      role: "user",
      content: "Classify this support ticket: I cannot log in after resetting my password.",
    },
  ],
  response_format: {
    type: "json_schema",
    json_schema: {
      type: "object",
      properties: {
        label: { type: "string", enum: ["billing", "auth", "bug", "other"] },
        confidence: { type: "number" },
        rationale: { type: "string" },
      },
      required: ["label", "confidence", "rationale"],
    },
  },
});

Schema failures are a normal runtime case. Workers AI can return JSON Mode couldn't be met when the requested schema cannot be satisfied, so handle that path with retry, fallback, or a human-review state. JSON Mode currently does not support streaming.

When To Add AI Gateway

Use the Workers AI binding directly when the app only needs Cloudflare-hosted models and simple request handling. Add AI Gateway when you need any of these:

• External providers such as OpenAI, Anthropic, Google AI Studio, Google Vertex AI, Azure OpenAI, Amazon Bedrock, Deepgram, or other supported providers.

• Provider observability: logs, analytics, custom metadata, and cost tracking.

• Cross-provider controls: caching, rate limiting, fallback, routing, or provider-level policy.

• A single place to compare or switch model providers without moving the rest of the Worker architecture.

Official Docs To Recheck

• Workers AI Models

• Workers AI Pricing

• Workers AI JSON Mode

• AI Gateway providers