{
  "provider_id": "anthropic",
  "vendor": "anthropic",
  "display_name": "Anthropic Claude",
  "_verified": "2026-10-10, for claude-sonnet-5-5 and claude-opus-5-5: tier ids, output caps (128,000) and context windows (1,000,000) read from the Models API (models.retrieve); both tiers listed and pinged clean via 'providers probe anthropic'. NOT repeated on the 5.5 ids: the over-limit 400 that confirmed the output cap, the request_constraints below, and the PDF measurements in file_handling. Those were established 2026-09-13 and 2026-09-24 on claude-sonnet-5 and claude-opus-5.",
  "tiers": {
    "light": {
      "model_id": "claude-sonnet-5-5",
      "display_name": "Claude Sonnet 5.5",
      "fallback_model_id": "",
      "api_key_env": "ANTHROPIC_API_KEY",
      "max_output_tokens": 128000,
      "max_context_tokens": 1000000,
      "cost_tier": "low",
      "thinking_levels": [
        "low",
        "medium",
        "high"
      ],
      "capabilities": [
        "text",
        "multimodal_image",
        "native_pdf",
        "structured_output",
        "function_calling"
      ],
      "_note": "One API key serves both tiers — Anthropic does not split keys by tier, unlike the two Gemini keys that isolate Flash volume from Pro quota. thinking_levels maps onto output_config.effort; the provider also accepts xhigh and max, which QueryHints has no vocabulary for. 'minimal' is absent because there is no such effort level."
    },
    "heavy": {
      "model_id": "claude-opus-5-5",
      "display_name": "Claude Opus 5.5",
      "fallback_model_id": "claude-sonnet-5-5",
      "_fallback_note": "Both are GA aliases rather than dated snapshots, so a retirement is unlikely; Sonnet 5.5 is named as the substitute because it shares the same request surface and caps.",
      "api_key_env": "ANTHROPIC_API_KEY",
      "max_output_tokens": 128000,
      "max_context_tokens": 1000000,
      "cost_tier": "high",
      "thinking_levels": [
        "low",
        "medium",
        "high"
      ],
      "capabilities": [
        "text",
        "multimodal_image",
        "native_pdf",
        "structured_output",
        "function_calling",
        "deep_reasoning"
      ]
    }
  },
  "request_constraints": {
    "_note": "Rejected parameters, verified against the Models API capability block. Sending any of these is a 400, so no client may pass them for this provider.",
    "thinking_budget_tokens": false,
    "sampling_params": false,
    "assistant_prefill": false,
    "thinking_mode": "adaptive",
    "stream_above_output_tokens": 32768
  },
  "file_handling": {
    "application/pdf": {
      "max_pages": 250,
      "max_size_mb": 24,
      "tokens_per_page": 2407,
      "tokens_per_page_by_resolution": {
        "low": 2407,
        "medium": 2407,
        "high": 2407
      },
      "default_media_resolution": "medium",
      "ingestion_paths": [
        "inline_bytes"
      ],
      "preferred_ingestion": "inline_bytes",
      "_verified": "MEASURED ON THE PREVIOUS MODELS, not on the 5.5 ids this manifest now names. 2026-09-24, both tiers (claude-sonnet-5, claude-opus-5), inline base64 document block. tokens_per_page MEASURED: (document-call input_tokens - same prompt as text) / pages on the 3-page probe PDF; identical on both tiers, and 60% above the 1500 placeholder it replaces. Page cost does not vary with media_resolution here, so the per-resolution block is filled uniformly. max_pages is the LARGEST COUNT VERIFIED, not a refusal: 101 and 250 pages were accepted and counted correctly; the 1,000,000-token context window binds at ~415 pages. max_size_mb is MEASURED against the decoded file: the documented 32 MB limit covers the request including base64, so a 20 MB PDF (27 MB encoded) was accepted and a 25 MB PDF (33 MB encoded) was refused twice. The refusal arrives as a dropped connection, which the client classifies as transient, so this guard must stop an oversized document before it is sent.",
      "_note": "Inline only, never the Files API: an upload is retained provider-side until deleted. No gs_uri path: this provider cannot read a GCS URI, so the dispatcher reads the bytes first (no remote_uri_gs capability)."
    }
  },
  "output_budget": {
    "_note": "Thinking tokens are output tokens and come out of max_tokens, as on Gemini — but budget_tokens was removed on the 5 family, so depth is set by output_config.effort and the spend is adaptive rather than capped. These reserves are what to hold back when sizing how much content one call may emit; max_tokens is a ceiling, not a charge, so reserving generously costs nothing unless spent.",
    "default_thinking_level": "high",
    "thinking_reserve_tokens": {
      "low": 4096,
      "medium": 16384,
      "high": 32768
    }
  },
  "document_strategies": {
    "application/pdf": "native",
    "application/json": "extract_text",
    "text/csv": "extract_text",
    "text/plain": "extract_text",
    "text/html": "extract_text",
    "default": "extract_and_chunk"
  }
}