{
  "version": 1,
  "entries": [
    {
      "id": "claude-subscription",
      "match": {
        "extension": "claude"
      },
      "policy": "refresh",
      "clock": "request",
      "tiers": [
        {
          "seconds": 300,
          "write": 1.25
        },
        {
          "seconds": 3600,
          "write": 2.0
        }
      ],
      "read": 0.1,
      "evidence": "documented",
      "source": "https://code.claude.com/docs/en/prompt-caching",
      "note": "Claude OAuth. Albedo's Claude requests use the 1h cache-write TTL for subscription and usage-credit paths; the provider still reports 5m writes for other clients. Lifetime counts from request start, so streaming time uses it up.",
      "checked": "2026-09-28"
    },
    {
      "id": "anthropic-api",
      "match": {
        "host": "api.anthropic.com"
      },
      "policy": "refresh",
      "clock": "request",
      "tiers": [
        {
          "seconds": 300,
          "write": 1.25
        },
        {
          "seconds": 3600,
          "write": 2.0
        }
      ],
      "read": 0.1,
      "evidence": "documented",
      "source": "https://platform.claude.com/docs/en/build-with-claude/prompt-caching",
      "note": "Every hit refreshes for free. 1h entries must precede 5m entries in one request. Minimum prefix is model-dependent (512-4096 tokens).",
      "checked": "2026-09-28"
    },
    {
      "id": "bedrock-claude",
      "match": {
        "host": "bedrock-runtime.*.amazonaws.com",
        "model": "*claude*"
      },
      "policy": "refresh",
      "clock": "request",
      "tiers": [
        {
          "seconds": 300,
          "write": 1.25
        },
        {
          "seconds": 3600,
          "write": 2.0
        }
      ],
      "read": 0.1,
      "evidence": "documented",
      "source": "https://docs.aws.amazon.com/bedrock/latest/userguide/prompt-caching.html",
      "note": "Resets with each successful hit. Cross-region inference can route to a cold region and force extra writes.",
      "checked": "2026-09-28"
    },
    {
      "id": "codex-subscription",
      "match": {
        "extension": "codex"
      },
      "policy": "evict",
      "survival": {
        "typical": 1800,
        "max": 86400
      },
      "evidence": "folklore",
      "source": "https://developers.openai.com/api/docs/guides/prompt-caching",
      "note": "ChatGPT/Codex OAuth: nothing published. Codex CLI sends only prompt_cache_key. Observed informally at about 30 minutes, which matches the API's GPT-5.6+ and 24h-retention behaviour.",
      "checked": "2026-09-28"
    },
    {
      "id": "openai-5.6",
      "match": {
        "host": "api.openai.com",
        "model": [
          "gpt-5.6*",
          "gpt-6*"
        ]
      },
      "policy": "refresh",
      "tiers": [
        {
          "seconds": 1800,
          "write": 1.25
        }
      ],
      "read": 0.1,
      "evidence": "documented",
      "source": "https://developers.openai.com/api/docs/guides/prompt-caching",
      "note": "prompt_cache_options.ttl \"30m\" is the only value and the default; a minimum, OpenAI may retain longer. Minimum prefix 1024 tokens.",
      "checked": "2026-09-28"
    },
    {
      "id": "openai",
      "match": {
        "host": "api.openai.com"
      },
      "policy": "evict",
      "survival": {
        "typical": 450,
        "max": 3600
      },
      "evidence": "documented",
      "source": "https://developers.openai.com/api/docs/guides/prompt-caching",
      "note": "Earlier models, prompt_cache_retention in_memory (albedo sends none): about 5-10 minutes idle, at most an hour. With \"24h\" retention: typically about 30 minutes, at most 24 hours.",
      "checked": "2026-09-28"
    },
    {
      "id": "gemini-implicit",
      "match": {
        "host": [
          "generativelanguage.googleapis.com",
          "*aiplatform.googleapis.com"
        ]
      },
      "policy": "evict",
      "survival": {
        "typical": 240,
        "max": 86400
      },
      "evidence": "folklore",
      "source": "https://ai.google.dev/gemini-api/docs/caching",
      "note": "Google publishes no implicit TTL. OpenRouter reports 3-5 minutes on average; Vertex says implicit entries are cleared within 24 hours.",
      "checked": "2026-09-28"
    },
    {
      "id": "antigravity",
      "match": {
        "extension": "antigravity"
      },
      "policy": "unknown",
      "survival": {
        "typical": 240,
        "max": 86400
      },
      "evidence": "unknown",
      "source": "https://ai.google.dev/gemini-api/docs/caching",
      "note": "Cloud Code Assist: no docs. Assumed to behave like Gemini implicit caching until measured.",
      "checked": "2026-09-28"
    },
    {
      "id": "deepseek",
      "match": {
        "host": "api.deepseek.com"
      },
      "policy": "evict",
      "survival": {
        "typical": 14400,
        "max": 259200
      },
      "evidence": "documented",
      "source": "https://api-docs.deepseek.com/guides/kv_cache",
      "note": "Disk cache, on by default; unused entries are cleared \"usually within a few hours to a few days\". Matches whole cache units only.",
      "checked": "2026-09-28"
    },
    {
      "id": "alibaba",
      "match": {
        "extension": "alibaba"
      },
      "policy": "evict",
      "survival": {
        "typical": 300,
        "max": 3600
      },
      "read": 0.2,
      "evidence": "documented",
      "source": "https://www.alibabacloud.com/help/en/model-studio/context-cache",
      "note": "Implicit cache (albedo sends no cache_control): validity \"indeterminate\", periodically cleared; hits 20%, writes 100%. Explicit markers would give a 5m refresh-on-hit clock at 1.25x write, 0.1x read. The typical and max are placeholders.",
      "checked": "2026-09-28"
    },
    {
      "id": "zai",
      "match": {
        "host": [
          "api.z.ai",
          "open.bigmodel.cn"
        ]
      },
      "policy": "unknown",
      "survival": {
        "typical": 300,
        "max": 3600
      },
      "evidence": "unknown",
      "source": "https://docs.z.ai/guides/capabilities/cache",
      "note": "Implicit only; docs give no validity period.",
      "checked": "2026-09-28"
    },
    {
      "id": "kimi-coding",
      "match": {
        "host": "api.kimi.com"
      },
      "policy": "evict",
      "survival": {
        "typical": 300,
        "max": 1800
      },
      "evidence": "measured",
      "source": "https://github.com/Leechael/pi-provider-kimi-code/blob/main/docs/caching.md",
      "note": "kimi-for-coding endpoint: probes hit at 60s and 300s and missed at 1800s and beyond; markers are ignored and caching works in 256-token chunks.",
      "checked": "2026-09-28"
    },
    {
      "id": "kimi-api",
      "match": {
        "host": [
          "api.moonshot.ai",
          "api.moonshot.cn"
        ]
      },
      "policy": "refresh",
      "tiers": [
        {
          "seconds": 300
        },
        {
          "seconds": 3600
        }
      ],
      "evidence": "documented",
      "source": "https://platform.kimi.ai/docs/guide/context-caching",
      "note": "5m unless prompt_cache_options.ttl asks for 1h; each hit renews under the TTL locked at first write. Organisation-scoped.",
      "checked": "2026-09-28"
    },
    {
      "id": "xai",
      "match": {
        "host": "api.x.ai"
      },
      "policy": "evict",
      "survival": {
        "typical": 300
      },
      "evidence": "documented",
      "source": "https://docs.x.ai/developers/advanced-api-usage/prompt-caching/best-practices",
      "note": "No TTL: entries \"can be evicted at any time due to server load or restarts\". x-grok-conv-id keeps requests on one cache. The typical is a placeholder.",
      "checked": "2026-09-28"
    },
    {
      "id": "groq",
      "match": {
        "host": "api.groq.com"
      },
      "policy": "refresh",
      "tiers": [
        {
          "seconds": 7200
        }
      ],
      "evidence": "documented",
      "source": "https://console.groq.com/docs/prompt-caching",
      "note": "Expires after 2 hours without use.",
      "checked": "2026-09-28"
    },
    {
      "id": "openrouter",
      "match": {
        "host": "openrouter.ai"
      },
      "policy": "unknown",
      "survival": {
        "typical": 300,
        "max": 3600
      },
      "evidence": "documented",
      "source": "https://openrouter.ai/docs/guides/best-practices/prompt-caching",
      "note": "No cache of its own: the upstream's rules apply, and TTL syntax is not translated between providers. Sticky routing to one upstream lapses after 10 minutes idle. Several reports of no Anthropic caching through it.",
      "checked": "2026-09-28"
    },
    {
      "id": "hyper",
      "match": {
        "host": "hyper.charm.land"
      },
      "policy": "unknown",
      "survival": {
        "typical": 300,
        "max": 3600
      },
      "evidence": "unknown",
      "source": "https://hyper.charm.land/docs/models.html",
      "note": "Gateway with no caching docs. Placeholder until measured.",
      "checked": "2026-09-28"
    }
  ]
}