Prompt caching

Prompt caching reuses the processed prefix of a long prompt, such as a system prompt, documents or tool definitions, across requests. Cache reads cost much less than normal input tokens. api14 passes the provider's cache pricing straight through: input_cache_read and input_cache_write are listed for each model in Models.

Automatic caching

These providers cache on their own. You don't need to change anything:

Put the parts that don't change at the start of the prompt and the parts that change at the end, so the prefix matches between requests.

Explicit caching with cache_control

Anthropic Claude (and Gemini explicit caching) need breakpoints. Add cache_control to a text content part:

{
  "model": "anthropic/claude-sonnet-4.5",
  "messages": [
    {
      "role": "system",
      "content": [
        { "type": "text", "text": "You are a contracts assistant." },
        {
          "type": "text",
          "text": "<the full 40-page agreement text>",
          "cache_control": { "type": "ephemeral" }
        }
      ]
    },
    { "role": "user", "content": "Which clause covers termination?" }
  ]
}

• Anthropic allows up to 4 breakpoints. The cache lives for about 5 minutes and is refreshed on each hit. Use "ttl": "1h" for a longer cache, which costs more to write.
• Cache writes on Anthropic cost about 1.25× input, and reads about 0.1×.
• To keep hitting the same provider's cache, add provider: { "order": ["Anthropic"], "allow_fallbacks": false } if you need it.

import OpenAI from "openai";

const client = new OpenAI({
  baseURL: "https://api14.com/api/v1",
  apiKey: process.env.API14_API_KEY,
});

const result = await client.chat.completions.create({
  "model": "anthropic/claude-sonnet-4.5",
  "messages": [
    {
      "role": "user",
      "content": "Explain vector databases in two sentences."
    }
  ]
});
console.log(result.choices[0].message.content);

Reading cache usage

The usage object shows cached tokens, and cost already includes the discount:

{
  "usage": {
    "prompt_tokens": 10450,
    "completion_tokens": 120,
    "total_tokens": 10570,
    "prompt_tokens_details": { "cached_tokens": 10200 },
    "cost": 0.004863
  }
}

cached_tokens also appears in your usage logs in the dashboard. Generation lookup shows the full breakdown, including cache_discount.

Errors

Caching never causes errors of its own. A malformed cache_control is ignored or rejected by the provider with a 400. See Errors.