Docs/LLM

Claude Opus 4.8

1M contextPrompt cachingThinking

POSThttps://api.sozdai.ai/v1/chat/completions

#Pricing Summary

Input Tokens

637,5 ₽ per 1M tokens. Corresponds to all context prompt tokens sent.

Output Tokens

3 187,5 ₽ per 1M tokens. Corresponds to content generated by the model.

Cache Read

63,75 ₽ per 1M tokens. Up to 90% savings when hitting the prompt cache.

Cache Write

796,88 ₽ per 1M tokens. Cost to index input context for caching.

#Example request

Execute completions using standard OpenAI SDK formats by pointing base endpoints here.

bash

curl https://api.sozdai.ai/v1/chat/completions \
  -H "Authorization: Bearer $CORRY_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "claude-opus-4-8",
    "messages": [{"role": "user", "content": "Hello"}]
  }'

#Response example

A standard OpenAI-compatible chat completion object. The model field is normalized to the name you requested.

json

{
  "id": "chatcmpl-abc123",
  "object": "chat.completion",
  "model": "claude-opus-4-8",
  "choices": [
    {
      "index": 0,
      "message": { "role": "assistant", "content": "Hello! How can I help you today?" },
      "finish_reason": "stop"
    }
  ],
  "usage": { "prompt_tokens": 12, "completion_tokens": 9, "total_tokens": 21, "cost": 0.000234, "currency": "USD" }
}

Cost field

usage.cost is the amount charged for this call (in USD); in streaming it arrives on the final chunk that carries usage. OpenAI SDKs ignore this extra field.

#Prompt caching

For cache-capable models, put the long static content (system, knowledge base) first and add a cache breakpoint. Cached portions can cost as little as 1/10. Cache hits are reported via the response usage field.

json

{
  "model": "claude-opus-4-8",
  "messages": [
    {
      "role": "system",
      "content": [
        { "type": "text", "text": "( long static context ... )", "cache_control": { "type": "ephemeral" } }
      ]
    },
    { "role": "user", "content": "your question" }
  ]
}

Prompt Caching Tip

Prompt caching matches request prefixes exactly. Make sure long contexts (e.g. system instructions, developer tools) appear at the start of your message list.

#Thinking

For models that support it, add reasoning_effort (low / medium / high). The thinking is returned in the response's reasoning_content field; thinking tokens are billed at the output rate.

json

{
  "model": "claude-opus-4-8",
  "max_tokens": 2048,
  "reasoning_effort": "high",
  "messages": [{ "role": "user", "content": "your question" }]
}