Prompt Caching

One option on the server handler cuts input token costs by ~90% for repeated system prompts. Zero client changes required.

How It Works

Anthropic's prompt caching lets you mark a prefix of your prompt for server-side caching. When the same prefix appears in subsequent requests, Anthropic serves it from cache at ~10% of the normal input token cost.

The system prompt is the perfect candidate — it's long, repeated on every request, and identical across turns.

cacheSystemPrompt: true detects the model provider at request time and automatically adds the cache_control breakpoint when an Anthropic model is used. No code changes needed per-request.

Setup

One line on the server handler:

// hooks.server.ts (SvelteKit) | app/api/ai/[...path]/route.ts (Next.js)
import { createStreamHandler } from "@aibind/sveltekit/server";

export const handle = createStreamHandler({
  models,
  cacheSystemPrompt: true, // ← that's it
});

Pass a system prompt from the client (or set a default on the server):

<script lang="ts">
  import { Stream } from "@aibind/sveltekit";

  const stream = new Stream({ model: "smart" });
</script>

<button
  onclick={() =>
    stream.send("What is the capital of France?", {
      system: "You are a helpful geography tutor. Answer concisely.",
    })}
>
  Ask
</button>

The first request caches the system prompt. Every subsequent request with the same system prompt costs ~90% less in input tokens.

Requirements

• Direct Anthropic provider — uses @ai-sdk/anthropic (anthropic(...)) as the model, not OpenRouter or a proxy. OpenRouter doesn't forward the cache_control metadata.
• System prompt present — caching only applies when a system field is included in the request.
• Anthropic model — the handler auto-detects the provider. Non-Anthropic models are unaffected and receive no cache_control.

import { anthropic } from "@ai-sdk/anthropic";
import { defineModels } from "@aibind/core";

// ✅ Works — direct @ai-sdk/anthropic
const models = defineModels({
  smart: anthropic("claude-sonnet-4-6"),
});

// ❌ Won't cache — OpenRouter proxies don't forward cache_control
const models = defineModels({
  smart: openrouter("anthropic/claude-sonnet-4-6"),
});

Mixing Models

You can mix Anthropic and non-Anthropic models freely. cacheSystemPrompt is a no-op for non-Anthropic models:

const models = defineModels({
  fast: openrouter("google/gemini-3.1-flash-lite-preview"), // unaffected
  smart: anthropic("claude-sonnet-4-6"), // cached
});

export const handle = createStreamHandler({
  models,
  cacheSystemPrompt: true,
});

Requests using fast go through normally. Requests using smart get the cache breakpoint added automatically.

API Reference

StreamHandlerConfig

Option	Type	Default	Description
cacheSystemPrompt	boolean	false	When true, adds cache_control: { type: "ephemeral" } to the system prompt for Anthropic models. Silently skipped for non-Anthropic providers.