Reduce OpenRouter costs with prompt caching (OpenAI automatic, Anthropic cache_control)

Documented steps

1. OpenAI models: caching is AUTOMATIC for prompts ≥ 1024 tokens — no request changes needed; cached reads are billed at a 0.25x–0.50x discount and writes cost nothing extra. Structure prompts with the static part first so prefixes match across calls.
2. Anthropic models: caching is EXPLICIT — add cache_control to a content block: {"role": "system", "content": [{"type": "text", "text": "<large static context>", "cache_control": {"type": "ephemeral"}}]}. Default TTL is 5 minutes; {"type": "ephemeral", "ttl": "1h"} extends to 1 hour.
3. Anthropic pricing: cache writes cost 1.25x normal input (2.0x for the 1h TTL); cache reads cost ~0.1x. Only cache blocks you will reuse within the TTL.
4. Other providers: DeepSeek caches automatically (reads ~0.1x, writes 1.0x); Grok caches automatically (reads 0.25x, no write premium); Gemini 2.5 has implicit caching.
5. Verify caching works: check response usage.prompt_tokens_details.cached_tokens, or fetch GET /api/v1/generation?id=<gen id> and read native_tokens_cached.
6. Docs: https://openrouter.ai/docs/guides/best-practices/prompt-caching