Language Models API

Overview

The Messages API is compatible with the official Anthropic API format. It supports streaming, vision, tool use, and extended thinking features.

Endpoint

Available Models

Prices are per 1M tokens

Request Parameters

Basic Request

curl -X POST https://api.apimodels.app/v1/messages \
  -H "Authorization: Bearer YOUR_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "claude-sonnet-4-5-20250929",
    "max_tokens": 1024,
    "messages": [
      {"role": "user", "content": "Hello, Claude!"}
    ]
  }'

Resposta

{
  "id": "msg_01XFDUDYJgAACzvnptvVoYEL",
  "type": "message",
  "role": "assistant",
  "content": [
    {
      "type": "text",
      "text": "Hello! How can I help you today?"
    }
  ],
  "model": "claude-sonnet-4-5-20250929",
  "stop_reason": "end_turn",
  "usage": {
    "input_tokens": 12,
    "output_tokens": 10
  }
}

Extended Thinking

For *-thinking models, you can enable extended thinking for complex reasoning tasks.

curl -X POST https://api.apimodels.app/v1/messages \
  -H "Authorization: Bearer YOUR_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "claude-sonnet-4-5-20250929-thinking",
    "max_tokens": 16000,
    "thinking": {
      "type": "enabled",
      "budget_tokens": 10000
    },
    "messages": [
      {"role": "user", "content": "Solve this step by step: What is 15% of 340?"}
    ]
  }'

Streaming

Set stream: true to receive Server-Sent Events (SSE). Strongly recommended for long generations: non-streaming requests that take longer than ~100 seconds to produce a first byte hit the CDN gateway timeout and return an HTML 524 page even though the generation completes (and is billed) server-side. If you expect more than a few hundred output tokens or use thinking/reasoning models, stream — or poll asynchronously.

curl -X POST https://api.apimodels.app/v1/messages \
  -H "Authorization: Bearer YOUR_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "claude-sonnet-4-5-20250929",
    "max_tokens": 1024,
    "stream": true,
    "messages": [
      {"role": "user", "content": "Write a short poem about coding."}
    ]
  }'

Prompt Caching

Repeated prefixes — a system prompt, a codebase dump, a long document — can be cached upstream and re-read at a fraction of the input rate. Cache prices are per model; see the price on each model page.

Claude: mark the last block of the stable prefix with cache_control. Reads come back as cache_read_input_tokens and writes as cache_creation_input_tokens, with the 5-minute and 1-hour TTLs broken out under cache_creation. Both TTLs bill at the same cache-write rate here, so passing ttl:"1h" costs nothing extra.

The cached prefix must be byte-identical between calls — one changed character (a timestamp, a re-ordered tool list) misses the cache and you pay the full input rate. Put everything variable at the END of the prompt.

{
  "model": "claude-fable-5-1",
  "max_tokens": 1024,
  "system": [
    {
      "type": "text",
      "text": "<long, unchanging instructions or document>",
      "cache_control": { "type": "ephemeral" }
    }
  ],
  "messages": [{ "role": "user", "content": "<the part that changes>" }]
}

Billing

Credits are calculated from actual token usage, across four tiers: full-rate input, cached input, cache writes and output. Reasoning / thinking tokens are billed at the OUTPUT rate and are not part of the visible answer — on Claude they are counted in output_tokens, on Gemini they arrive as thoughtsTokenCount alongside candidatesTokenCount, and on GPT-5.6 as reasoning_tokens inside completion_tokens. A model that thinks a lot can bill several times what the visible reply suggests. Failed requests are never billed.

Cost = (input_tokens * input_price + cached_tokens * cache_read_price
      + cache_write_tokens * cache_write_price + output_tokens * output_price) / 1,000,000
Credits = Cost  // amounts are in USD ($)

Error Codes