feat(ocr): add Reducto parse OCR support (#26068)

* feat(ocr): add Reducto parse OCR support

* fix(reducto): address OCR review feedback

* chore: refresh uv lockfile

* Revert "chore: refresh uv lockfile"

This reverts commit 47200c0e60.

docs/my-website/docs/providers/reducto.md

@@ -0,0 +1,103 @@
+# Reducto
+
+## Overview
+
+| Property | Details |
+|-------|-------|
+| Description | Reducto parse support over LiteLLM's existing OCR API |
+| Provider Route on LiteLLM | `reducto/` |
+| Supported Operations | `/ocr` |
+| Supported Models | `reducto/parse-v3`, `reducto/parse-legacy` |
+| Link to Provider Doc | [Reducto ↗](https://platform.reducto.ai/) |
+
+Reducto is exposed through LiteLLM's OCR surface, so this provider uses `litellm.ocr()` and `litellm.aocr()`.
+
+## Quick Start
+
+### LiteLLM SDK
+
+```python showLineNumbers title="SDK Usage"
+import litellm
+import os
+
+os.environ["REDUCTO_API_KEY"] = "your-api-key"
+
+response = litellm.ocr(
+    model="reducto/parse-v3",
+    document={"type": "file", "file": "document.pdf"},
+)
+
+for page in response.pages:
+    print(page.markdown)
+```
+
+You can also override credentials per call with `api_key=` and `api_base=`.
+
+### LiteLLM Proxy
+
+```yaml showLineNumbers title="proxy_config.yaml"
+model_list:
+  - model_name: reducto-parse
+    litellm_params:
+      model: reducto/parse-v3
+      api_key: os.environ/REDUCTO_API_KEY
+    model_info:
+      mode: ocr
+```
+
+## Parse V3
+
+`reducto/parse-v3` maps to Reducto's current parse API and accepts:
+
+- `formatting`
+- `retrieval`
+- `settings`
+
+```python showLineNumbers title="Parse V3"
+response = await litellm.aocr(
+    model="reducto/parse-v3",
+    document={"type": "file", "file": "document.pdf"},
+    formatting={"table_output_format": "html"},
+    retrieval={"chunk_mode": "section"},
+    settings={"ocr_system": "standard"},
+)
+```
+
+## Parse Legacy
+
+`reducto/parse-legacy` keeps the legacy request shape and accepts `enhance`.
+
+```python showLineNumbers title="Parse Legacy"
+response = litellm.ocr(
+    model="reducto/parse-legacy",
+    document={"type": "file", "file": "document.pdf"},
+    enhance={"agentic": [{"type": "table"}]},
+)
+```
+
+## Upload Behavior
+
+- `document={"type":"file","file":...}` is auto-converted by LiteLLM into a data URI, then uploaded to Reducto's `/upload` endpoint before `/parse`.
+- `document_url="reducto://..."` is passed through directly and skips upload.
+- Plain `http(s)` document URLs are rejected for Reducto. Upload the file first or pass a local file to LiteLLM.
+- Image files also work through `type="file"`; LiteLLM normalizes them to `image_url` data URIs before the Reducto upload step.
+
+## Cost Tracking
+
+Reducto returns OCR usage in credits. LiteLLM supports credit-priced OCR models via `ocr_cost_per_credit`.
+
+If you want spend tracking, register pricing for your deployment:
+
+```python showLineNumbers title="Register OCR Credit Pricing"
+import litellm
+
+litellm.register_model(
+    {
+        "reducto/parse-v3": {
+            "litellm_provider": "reducto",
+            "mode": "ocr",
+            "ocr_cost_per_credit": 0.003,
+        }
+    }
+)
+```

docs/my-website/sidebars.js

@@ -996,6 +996,7 @@ const sidebars = {
         "providers/predibase",
         "providers/pydantic_ai_agent",
         "providers/ragflow",
+        "providers/reducto",
         "providers/recraft",
         "providers/replicate",
         {

litellm/cost_calculator.py

@@ -1801,10 +1801,6 @@ def ocr_cost(
     if response.usage_info is None:
         raise ValueError("OCR response usage_info is None")
 
-    pages_processed = response.usage_info.pages_processed
-    if pages_processed is None:
-        raise ValueError("OCR response pages_processed is None")
-
     try:
         model_info: Optional[ModelInfo] = litellm.get_model_info(
             model=model, custom_llm_provider=custom_llm_provider
@@ -1812,6 +1808,17 @@ def ocr_cost(
     except Exception:
         model_info = None
 
+    credits = getattr(response.usage_info, "credits", None)
+    cost_per_credit = None
+    if model_info is not None:
+        cost_per_credit = model_info.get("ocr_cost_per_credit")
+    if credits is not None and cost_per_credit:
+        return cost_per_credit * credits, 0.0
+
+    pages_processed = response.usage_info.pages_processed
+    if pages_processed is None:
+        return 0.0, 0.0
+
     ocr_cost_per_page: float = 0.0
     if model_info is not None:
         ocr_cost_per_page = model_info.get("ocr_cost_per_page") or 0.0

litellm/llms/base_llm/ocr/transformation.py

@@ -54,6 +54,7 @@ class OCRUsageInfo(LiteLLMPydanticObjectBase):
     """Usage information from OCR response."""
 
     pages_processed: Optional[int] = None
+    credits: Optional[float] = None
     doc_size_bytes: Optional[int] = None
 
     model_config = {"extra": "allow"}

litellm/llms/reducto/__init__.py

@@ -0,0 +1 @@
+

litellm/llms/reducto/common.py

@@ -0,0 +1,141 @@
+import base64
+import binascii
+from collections import defaultdict
+from typing import TYPE_CHECKING, Any, Dict, List, NoReturn, Optional, Tuple
+
+from litellm.constants import request_timeout
+
+REDUCTO_API_BASE = "https://platform.reducto.ai"
+REDUCTO_ID_PREFIX = "reducto://"
+
+if TYPE_CHECKING:
+    from litellm.llms.base_llm.ocr.transformation import OCRPage
+
+
+def _normalize_api_base(api_base: Optional[str]) -> str:
+    return (api_base or REDUCTO_API_BASE).rstrip("/")
+
+
+def _raise_bad_request(message: str, model: str) -> NoReturn:
+    import litellm
+
+    raise litellm.BadRequestError(
+        message=message,
+        model=model,
+        llm_provider="reducto",
+    )
+
+
+def extract_file_id_or_bytes(
+    source_url: str,
+    model: str,
+) -> Tuple[Optional[str], Optional[bytes], Optional[str]]:
+    if source_url.startswith(REDUCTO_ID_PREFIX):
+        return source_url, None, None
+
+    if source_url.startswith("http://") or source_url.startswith("https://"):
+        _raise_bad_request(
+            "Reducto requires type='file' (auto-uploaded) or a reducto:// id. Plain http(s) URLs are not supported; upload the file first.",
+            model=model,
+        )
+
+    if not source_url.startswith("data:"):
+        _raise_bad_request(
+            "Reducto requires a reducto:// id or a base64 data URI after OCR preprocessing.",
+            model=model,
+        )
+
+    try:
+        header, encoded = source_url.split(",", 1)
+    except ValueError:
+        _raise_bad_request("Invalid Reducto data URI provided.", model=model)
+
+    if ";base64" not in header:
+        _raise_bad_request(
+            "Reducto only supports base64-encoded data URIs.", model=model
+        )
+
+    mime = header.removeprefix("data:").split(";")[0] or "application/octet-stream"
+    try:
+        raw_bytes = base64.b64decode(encoded, validate=True)
+    except (binascii.Error, ValueError):
+        _raise_bad_request("Invalid Reducto base64 payload provided.", model=model)
+
+    return None, raw_bytes, mime
+
+
+def upload_bytes_sync(
+    raw_bytes: bytes,
+    mime: Optional[str],
+    api_key: str,
+    api_base: Optional[str],
+) -> str:
+    import litellm
+
+    response = litellm.module_level_client.post(
+        url="{}{}".format(_normalize_api_base(api_base), "/upload"),
+        headers={"Authorization": f"Bearer {api_key}"},
+        files={"file": ("document", raw_bytes, mime or "application/octet-stream")},
+        timeout=request_timeout,
+    )
+    response.raise_for_status()
+    return response.json()["file_id"]
+
+
+async def upload_bytes_async(
+    raw_bytes: bytes,
+    mime: Optional[str],
+    api_key: str,
+    api_base: Optional[str],
+) -> str:
+    import litellm
+
+    response = await litellm.module_level_aclient.post(
+        url="{}{}".format(_normalize_api_base(api_base), "/upload"),
+        headers={"Authorization": f"Bearer {api_key}"},
+        files={"file": ("document", raw_bytes, mime or "application/octet-stream")},
+        timeout=request_timeout,
+    )
+    response.raise_for_status()
+    return response.json()["file_id"]
+
+
+def build_pages_from_reducto(result: Dict[str, Any]) -> List["OCRPage"]:
+    from litellm.llms.base_llm.ocr.transformation import OCRPage
+
+    chunks = result.get("chunks", []) or []
+    blocks_by_page: Dict[int, List[Dict[str, Any]]] = defaultdict(list)
+
+    for chunk in chunks:
+        for block in chunk.get("blocks", []) or []:
+            page_no = (block.get("bbox") or {}).get("page")
+            if page_no is None:
+                continue
+            try:
+                normalized_page = int(page_no)
+            except (TypeError, ValueError):
+                continue
+            blocks_by_page[normalized_page].append(block)
+
+    if not blocks_by_page:
+        fallback_markdown = "\n\n".join(
+            chunk.get("content", "") for chunk in chunks if chunk.get("content")
+        )
+        if fallback_markdown == "":
+            return []
+        return [OCRPage(index=0, markdown=fallback_markdown)]
+
+    pages: List["OCRPage"] = []
+    for page_no, blocks in sorted(blocks_by_page.items()):
+        markdown = "\n\n".join(
+            block.get("content", "") for block in blocks if block.get("content")
+        )
+        page_index = max(page_no - 1, 0)
+        pages.append(
+            OCRPage(
+                index=page_index,
+                markdown=markdown,
+                blocks=blocks,
+            )
+        )
+    return pages

litellm/llms/reducto/ocr/__init__.py

@@ -0,0 +1 @@
+

litellm/llms/reducto/ocr/transformation.py

@@ -0,0 +1,205 @@
+from typing import Any, Dict, Optional
+
+import httpx
+
+from litellm.llms.base_llm.ocr.transformation import (
+    BaseOCRConfig,
+    DocumentType,
+    OCRRequestData,
+    OCRResponse,
+    OCRUsageInfo,
+)
+from litellm.llms.reducto.common import (
+    REDUCTO_API_BASE,
+    build_pages_from_reducto,
+    extract_file_id_or_bytes,
+    upload_bytes_async,
+    upload_bytes_sync,
+)
+
+
+class _BaseReductoOCRConfig(BaseOCRConfig):
+    def __init__(self) -> None:
+        super().__init__()
+        self._api_key: Optional[str] = None
+        self._api_base: Optional[str] = None
+
+    def map_ocr_params(
+        self,
+        non_default_params: dict,
+        optional_params: dict,
+        model: str,
+    ) -> dict:
+        mapped_params = dict(optional_params)
+        supported_params = self.get_supported_ocr_params(model=model)
+        for param, value in non_default_params.items():
+            if param in supported_params:
+                mapped_params[param] = value
+        return mapped_params
+
+    def validate_environment(
+        self,
+        headers: Dict,
+        model: str,
+        api_key: Optional[str] = None,
+        api_base: Optional[str] = None,
+        litellm_params: Optional[dict] = None,
+        **kwargs,
+    ) -> Dict:
+        from litellm.secret_managers.main import get_secret_str
+
+        resolved_key = api_key or get_secret_str("REDUCTO_API_KEY")
+        if resolved_key is None:
+            raise ValueError(
+                "Missing REDUCTO_API_KEY - set it in the environment or pass api_key to litellm.ocr()/litellm.aocr()"
+            )
+
+        self._api_key = resolved_key
+        self._api_base = (api_base or REDUCTO_API_BASE).rstrip("/")
+
+        return {
+            "Authorization": f"Bearer {resolved_key}",
+            "Content-Type": "application/json",
+            **headers,
+        }
+
+    def get_complete_url(
+        self,
+        api_base: Optional[str],
+        model: str,
+        optional_params: dict,
+        litellm_params: Optional[dict] = None,
+        **kwargs,
+    ) -> str:
+        return "{}/parse".format((api_base or REDUCTO_API_BASE).rstrip("/"))
+
+    def _get_source_url(self, document: DocumentType, model: str) -> str:
+        source_url = document.get("document_url") or document.get("image_url")
+        if source_url is None:
+            raise ValueError(
+                "Reducto expected OCR preprocessing to produce document_url or image_url for model={}".format(
+                    model
+                )
+            )
+        return source_url
+
+    def _ensure_file_id_sync(self, model: str, document: DocumentType) -> str:
+        source_url = self._get_source_url(document=document, model=model)
+        file_id, raw_bytes, mime = extract_file_id_or_bytes(source_url, model=model)
+        if file_id is not None:
+            return file_id
+        if self._api_key is None:
+            raise ValueError("Reducto API key was not initialized before OCR upload.")
+        return upload_bytes_sync(
+            raw_bytes=raw_bytes or b"",
+            mime=mime,
+            api_key=self._api_key,
+            api_base=self._api_base,
+        )
+
+    async def _ensure_file_id_async(self, model: str, document: DocumentType) -> str:
+        source_url = self._get_source_url(document=document, model=model)
+        file_id, raw_bytes, mime = extract_file_id_or_bytes(source_url, model=model)
+        if file_id is not None:
+            return file_id
+        if self._api_key is None:
+            raise ValueError("Reducto API key was not initialized before OCR upload.")
+        return await upload_bytes_async(
+            raw_bytes=raw_bytes or b"",
+            mime=mime,
+            api_key=self._api_key,
+            api_base=self._api_base,
+        )
+
+    def transform_ocr_response(
+        self,
+        model: str,
+        raw_response: httpx.Response,
+        logging_obj: Any,
+        **kwargs,
+    ) -> OCRResponse:
+        response_json = raw_response.json()
+        result = response_json.get("result", response_json) or {}
+        usage = response_json.get("usage", {}) or {}
+        response = OCRResponse(
+            pages=build_pages_from_reducto(result),
+            model=model,
+            usage_info=OCRUsageInfo(
+                pages_processed=usage.get("num_pages"),
+                credits=usage.get("credits"),
+            ),
+            object="ocr",
+        )
+        response._hidden_params["reducto_raw"] = response_json
+        return response
+
+
+class ReductoParseV3Config(_BaseReductoOCRConfig):
+    def get_supported_ocr_params(self, model: str) -> list:
+        return ["formatting", "retrieval", "settings"]
+
+    def transform_ocr_request(
+        self,
+        model: str,
+        document: DocumentType,
+        optional_params: dict,
+        headers: dict,
+        **kwargs,
+    ) -> OCRRequestData:
+        file_id = self._ensure_file_id_sync(model=model, document=document)
+        return OCRRequestData(data={"input": file_id, **optional_params}, files=None)
+
+    async def async_transform_ocr_request(
+        self,
+        model: str,
+        document: DocumentType,
+        optional_params: dict,
+        headers: dict,
+        **kwargs,
+    ) -> OCRRequestData:
+        file_id = await self._ensure_file_id_async(model=model, document=document)
+        return OCRRequestData(data={"input": file_id, **optional_params}, files=None)
+
+
+class ReductoParseLegacyConfig(_BaseReductoOCRConfig):
+    def get_supported_ocr_params(self, model: str) -> list:
+        return ["enhance"]
+
+    def _build_legacy_body(self, file_id: str, optional_params: dict) -> Dict[str, Any]:
+        body: Dict[str, Any] = {"document_url": file_id}
+        enhance = optional_params.get("enhance")
+        if enhance is not None:
+            body["options"] = {"enhance": enhance}
+        return body
+
+    def transform_ocr_request(
+        self,
+        model: str,
+        document: DocumentType,
+        optional_params: dict,
+        headers: dict,
+        **kwargs,
+    ) -> OCRRequestData:
+        file_id = self._ensure_file_id_sync(model=model, document=document)
+        return OCRRequestData(
+            data=self._build_legacy_body(
+                file_id=file_id, optional_params=optional_params
+            ),
+            files=None,
+        )
+
+    async def async_transform_ocr_request(
+        self,
+        model: str,
+        document: DocumentType,
+        optional_params: dict,
+        headers: dict,
+        **kwargs,
+    ) -> OCRRequestData:
+        file_id = await self._ensure_file_id_async(model=model, document=document)
+        return OCRRequestData(
+            data=self._build_legacy_body(
+                file_id=file_id, optional_params=optional_params
+            ),
+            files=None,
+        )

litellm/model_prices_and_context_window_backup.json

@@ -27379,6 +27379,20 @@
         "supports_tool_choice": true,
         "source": "https://aws.amazon.com/bedrock/pricing/"
     },
+    "reducto/parse-legacy": {
+        "litellm_provider": "reducto",
+        "mode": "ocr",
+        "supported_endpoints": [
+            "/v1/ocr"
+        ]
+    },
+    "reducto/parse-v3": {
+        "litellm_provider": "reducto",
+        "mode": "ocr",
+        "supported_endpoints": [
+            "/v1/ocr"
+        ]
+    },
     "recraft/recraftv2": {
         "litellm_provider": "recraft",
         "mode": "image_generation",

litellm/types/utils.py

@@ -239,6 +239,7 @@ class ModelInfoBase(ProviderSpecificModelInfo, total=False):
         float
     ]  # video_generation tier: key output_cost_per_second_<resolution> (e.g. 1080p, 720p)
     ocr_cost_per_page: Optional[float]  # for OCR models
+    ocr_cost_per_credit: Optional[float]  # for OCR models priced by credit
     annotation_cost_per_page: Optional[float]  # for OCR models
     search_context_cost_per_query: Optional[
         SearchContextCostPerQuery
@@ -256,6 +257,7 @@ class ModelInfoBase(ProviderSpecificModelInfo, total=False):
             "chat",
             "audio_transcription",
             "responses",
+            "ocr",
         ]
     ]
     tpm: Optional[int]
@@ -3191,6 +3193,7 @@ class LlmProviders(str, Enum):
     ANTHROPIC_TEXT = "anthropic_text"
     BYTEZ = "bytez"
     REPLICATE = "replicate"
+    REDUCTO = "reducto"
     RUNWAYML = "runwayml"
     AWS_POLLY = "aws_polly"
     HUGGINGFACE = "huggingface"

litellm/utils.py

@@ -5909,6 +5909,7 @@ def _get_model_info_helper(  # noqa: PLR0915
                 tpm=_model_info.get("tpm", None),
                 rpm=_model_info.get("rpm", None),
                 ocr_cost_per_page=_model_info.get("ocr_cost_per_page", None),
+                ocr_cost_per_credit=_model_info.get("ocr_cost_per_credit", None),
                 annotation_cost_per_page=_model_info.get(
                     "annotation_cost_per_page", None
                 ),
@@ -9128,6 +9129,18 @@ class ProviderConfigManager:
 
             return get_vertex_ai_ocr_config(model=model)
 
+        if provider == litellm.LlmProviders.REDUCTO:
+            from litellm.llms.reducto.ocr.transformation import (
+                ReductoParseLegacyConfig,
+                ReductoParseV3Config,
+            )
+
+            if model == "parse-v3":
+                return ReductoParseV3Config()
+            if model == "parse-legacy":
+                return ReductoParseLegacyConfig()
+            return None
+
         MistralOCRConfig = getattr(sys.modules[__name__], "MistralOCRConfig")
         PROVIDER_TO_CONFIG_MAP = {
             litellm.LlmProviders.MISTRAL: MistralOCRConfig,

tests/proxy_unit_tests/test_reducto_ocr_route.py

@@ -0,0 +1,82 @@
+import asyncio
+import os
+from unittest.mock import AsyncMock, patch
+
+import litellm
+import pytest
+from fastapi.testclient import TestClient
+
+from litellm.llms.base_llm.ocr.transformation import OCRPage, OCRResponse, OCRUsageInfo
+from litellm.proxy.proxy_server import app, initialize
+
+
+@pytest.fixture(scope="function")
+def fake_env_vars(monkeypatch):
+    monkeypatch.setenv("OPENAI_API_KEY", "fake_openai_api_key")
+    monkeypatch.setenv("OPENAI_API_BASE", "http://fake-openai-api-base")
+    monkeypatch.setenv("AZURE_AI_API_BASE", "http://fake-azure-api-base")
+    monkeypatch.setenv("AZURE_AI_API_KEY", "fake_azure_api_key")
+    monkeypatch.setenv("AZURE_OPENAI_API_KEY", "fake_azure_openai_api_key")
+    monkeypatch.setenv("AZURE_SWEDEN_API_BASE", "http://fake-azure-sweden-api-base")
+    monkeypatch.setenv("AZURE_SWEDEN_API_KEY", "fake_azure_sweden_api_key")
+    monkeypatch.setenv("REDIS_HOST", "localhost")
+
+
+@pytest.fixture(scope="function")
+def client_no_auth(fake_env_vars):
+    from litellm.proxy.proxy_server import cleanup_router_config_variables
+
+    original_disable_aiohttp = litellm.disable_aiohttp_transport
+    litellm.disable_aiohttp_transport = True
+    litellm.in_memory_llm_clients_cache.flush_cache()
+    cleanup_router_config_variables()
+
+    filepath = os.path.dirname(os.path.abspath(__file__))
+    config_fp = os.path.join(filepath, "test_configs", "test_config_no_auth.yaml")
+    asyncio.run(initialize(config=config_fp, debug=True))
+
+    try:
+        yield TestClient(app)
+    finally:
+        litellm.disable_aiohttp_transport = original_disable_aiohttp
+        litellm.in_memory_llm_clients_cache.flush_cache()
+
+
+def test_proxy_reducto_ocr_json_passthrough(client_no_auth):
+    mocked_response = OCRResponse(
+        pages=[OCRPage(index=0, markdown="Proxy OCR")],
+        model="parse-v3",
+        usage_info=OCRUsageInfo(pages_processed=1, credits=1),
+    )
+
+    with patch(
+        "litellm.proxy.proxy_server.llm_router.aocr",
+        new=AsyncMock(return_value=mocked_response),
+    ) as mock_aocr:
+        response = client_no_auth.post(
+            "/v1/ocr",
+            json={
+                "model": "reducto/parse-v3",
+                "document": {
+                    "type": "document_url",
+                    "document_url": "reducto://proxy.pdf",
+                },
+                "api_key": "proxy-key",
+                "api_base": "https://platform.reducto.ai",
+            },
+        )
+
+    assert response.status_code == 200
+    assert mock_aocr.await_count == 1
+    assert mock_aocr.await_args.kwargs["model"] == "reducto/parse-v3"
+    assert mock_aocr.await_args.kwargs["document"] == {
+        "type": "document_url",
+        "document_url": "reducto://proxy.pdf",
+    }
+    assert mock_aocr.await_args.kwargs["api_key"] == "proxy-key"
+    assert mock_aocr.await_args.kwargs["api_base"] == "https://platform.reducto.ai"
+
+    response_body = response.json()
+    assert response_body["object"] == "ocr"
+    assert response_body["usage_info"]["credits"] == 1
+    assert response_body["pages"][0]["markdown"] == "Proxy OCR"

tests/test_litellm/llms/reducto/__init__.py

@@ -0,0 +1 @@
+

tests/test_litellm/llms/reducto/test_cost.py

@@ -0,0 +1,73 @@
+import litellm
+
+from litellm.cost_calculator import completion_cost
+from litellm.llms.base_llm.ocr.transformation import OCRPage, OCRResponse, OCRUsageInfo
+
+
+def test_ocr_cost_prefers_credit_pricing_when_pages_processed_is_none(monkeypatch):
+    monkeypatch.setattr(
+        litellm,
+        "get_model_info",
+        lambda model, custom_llm_provider=None: {"ocr_cost_per_credit": 0.003},
+    )
+
+    response = OCRResponse(
+        pages=[OCRPage(index=0, markdown="credit priced")],
+        model="parse-v3",
+        usage_info=OCRUsageInfo(pages_processed=None, credits=10),
+    )
+
+    cost = completion_cost(
+        completion_response=response,
+        model="reducto/parse-v3",
+        custom_llm_provider="reducto",
+        call_type="ocr",
+    )
+
+    assert cost == 0.03
+
+
+def test_ocr_cost_falls_back_to_page_pricing(monkeypatch):
+    monkeypatch.setattr(
+        litellm,
+        "get_model_info",
+        lambda model, custom_llm_provider=None: {"ocr_cost_per_page": 0.5},
+    )
+
+    response = OCRResponse(
+        pages=[OCRPage(index=0, markdown="page priced")],
+        model="mistral-ocr-latest",
+        usage_info=OCRUsageInfo(pages_processed=2),
+    )
+
+    cost = completion_cost(
+        completion_response=response,
+        model="mistral/mistral-ocr-latest",
+        custom_llm_provider="mistral",
+        call_type="ocr",
+    )
+
+    assert cost == 1.0
+
+
+def test_ocr_cost_returns_zero_when_no_pricing_and_no_pages(monkeypatch):
+    monkeypatch.setattr(
+        litellm,
+        "get_model_info",
+        lambda model, custom_llm_provider=None: {},
+    )
+
+    response = OCRResponse(
+        pages=[OCRPage(index=0, markdown="unpriced")],
+        model="parse-v3",
+        usage_info=OCRUsageInfo(pages_processed=None, credits=5),
+    )
+
+    cost = completion_cost(
+        completion_response=response,
+        model="reducto/parse-v3",
+        custom_llm_provider="reducto",
+        call_type="ocr",
+    )
+
+    assert cost == 0.0

tests/test_litellm/llms/reducto/test_model_info.py

@@ -0,0 +1,44 @@
+import uuid
+
+import litellm
+
+from litellm.utils import _invalidate_model_cost_lowercase_map
+
+
+def test_reducto_provider_registration():
+    model, custom_llm_provider, _, _ = litellm.get_llm_provider(
+        model="reducto/parse-v3"
+    )
+
+    assert model == "parse-v3"
+    assert custom_llm_provider == "reducto"
+
+
+def test_get_model_info_preserves_ocr_cost_per_credit():
+    test_model_name = f"reducto/test-cost-propagation-{uuid.uuid4().hex[:12]}"
+    previous_model_entry = litellm.model_cost.get(test_model_name)
+    _invalidate_model_cost_lowercase_map()
+
+    try:
+        litellm.register_model(
+            {
+                test_model_name: {
+                    "litellm_provider": "reducto",
+                    "mode": "ocr",
+                    "ocr_cost_per_credit": 0.003,
+                }
+            }
+        )
+
+        model_info = litellm.get_model_info(
+            model=test_model_name,
+            custom_llm_provider="reducto",
+        )
+
+        assert model_info.get("ocr_cost_per_credit") == 0.003
+    finally:
+        if previous_model_entry is None:
+            litellm.model_cost.pop(test_model_name, None)
+        else:
+            litellm.model_cost[test_model_name] = previous_model_entry
+        _invalidate_model_cost_lowercase_map()

tests/test_litellm/llms/reducto/test_parse_legacy.py

@@ -0,0 +1,59 @@
+import json
+
+import litellm
+import pytest
+
+
+@pytest.fixture()
+def disable_aiohttp_transport():
+    original_disable_aiohttp = litellm.disable_aiohttp_transport
+    litellm.disable_aiohttp_transport = True
+    litellm.in_memory_llm_clients_cache.flush_cache()
+    try:
+        yield
+    finally:
+        litellm.disable_aiohttp_transport = original_disable_aiohttp
+        litellm.in_memory_llm_clients_cache.flush_cache()
+
+
+@pytest.mark.asyncio
+async def test_parse_legacy_wraps_enhance_under_options(
+    disable_aiohttp_transport, respx_mock
+):
+    upload_route = respx_mock.post("https://platform.reducto.ai/upload").respond(
+        json={"file_id": "reducto://legacy.pdf"}
+    )
+    parse_route = respx_mock.post("https://platform.reducto.ai/parse").respond(
+        json={
+            "usage": {"num_pages": 1, "credits": 1},
+            "result": {
+                "chunks": [
+                    {
+                        "content": "Legacy parse",
+                        "blocks": [{"content": "Legacy parse", "bbox": {"page": 1}}],
+                    }
+                ]
+            },
+        }
+    )
+
+    response = await litellm.aocr(
+        model="reducto/parse-legacy",
+        document={
+            "type": "file",
+            "file": b"%PDF-1.4 legacy",
+            "mime_type": "application/pdf",
+        },
+        api_key="legacy-key",
+        api_base="https://platform.reducto.ai",
+        enhance={"agentic": [{"type": "table"}]},
+    )
+
+    assert upload_route.called
+    assert parse_route.called
+    request_body = json.loads(parse_route.calls[0].request.read())
+    assert request_body == {
+        "document_url": "reducto://legacy.pdf",
+        "options": {"enhance": {"agentic": [{"type": "table"}]}},
+    }
+    assert response.pages[0].markdown == "Legacy parse"

tests/test_litellm/llms/reducto/test_parse_v3.py

@@ -0,0 +1,152 @@
+import json
+
+import litellm
+import pytest
+
+
+def _reducto_parse_response() -> dict:
+    return {
+        "job_id": "job_123",
+        "usage": {"num_pages": 3, "credits": 3},
+        "result": {
+            "chunks": [
+                {
+                    "content": "Page 1 block A",
+                    "blocks": [
+                        {
+                            "content": "Page 1 block A",
+                            "bbox": {"page": 1},
+                            "kind": "text",
+                        }
+                    ],
+                },
+                {
+                    "content": "Page 2 block A",
+                    "blocks": [
+                        {
+                            "content": "Page 2 block A",
+                            "bbox": {"page": 2},
+                            "kind": "table",
+                        }
+                    ],
+                },
+                {
+                    "content": "Page 1 block B",
+                    "blocks": [
+                        {
+                            "content": "Page 1 block B",
+                            "bbox": {"page": 1},
+                            "kind": "text",
+                        }
+                    ],
+                },
+                {
+                    "content": "Page 3 block A",
+                    "blocks": [
+                        {
+                            "content": "Page 3 block A",
+                            "bbox": {"page": 3},
+                            "kind": "figure",
+                        }
+                    ],
+                },
+            ]
+        },
+    }
+
+
+@pytest.fixture()
+def disable_aiohttp_transport():
+    original_disable_aiohttp = litellm.disable_aiohttp_transport
+    litellm.disable_aiohttp_transport = True
+    litellm.in_memory_llm_clients_cache.flush_cache()
+    try:
+        yield
+    finally:
+        litellm.disable_aiohttp_transport = original_disable_aiohttp
+        litellm.in_memory_llm_clients_cache.flush_cache()
+
+
+@pytest.mark.asyncio
+async def test_parse_v3_file_upload_and_response_mapping(
+    disable_aiohttp_transport, respx_mock
+):
+    upload_route = respx_mock.post("https://platform.reducto.ai/upload").respond(
+        json={"file_id": "reducto://uploaded.pdf"}
+    )
+    parse_route = respx_mock.post("https://platform.reducto.ai/parse").respond(
+        json=_reducto_parse_response()
+    )
+
+    response = await litellm.aocr(
+        model="reducto/parse-v3",
+        document={
+            "type": "file",
+            "file": b"%PDF-1.4 reducto",
+            "mime_type": "application/pdf",
+        },
+        api_key="test-key",
+        api_base="https://platform.reducto.ai",
+        formatting={"table_output_format": "html"},
+        retrieval={"chunk_mode": "section"},
+        settings={"ocr_system": "standard"},
+    )
+
+    assert upload_route.called
+    assert parse_route.called
+    assert len(upload_route.calls) == 1
+    assert len(parse_route.calls) == 1
+
+    upload_request = upload_route.calls[0].request
+    assert upload_request.headers["authorization"] == "Bearer test-key"
+    assert "application/json" not in upload_request.headers["content-type"]
+    upload_body = upload_request.read()
+    assert b'filename="document"' in upload_body
+    assert b"application/pdf" in upload_body
+
+    parse_request_body = json.loads(parse_route.calls[0].request.read())
+    assert parse_request_body["input"] == "reducto://uploaded.pdf"
+    assert parse_request_body["formatting"] == {"table_output_format": "html"}
+    assert parse_request_body["retrieval"] == {"chunk_mode": "section"}
+    assert parse_request_body["settings"] == {"ocr_system": "standard"}
+
+    assert response.usage_info is not None
+    assert response.usage_info.credits == 3
+    assert response.usage_info.pages_processed == 3
+    assert len(response.pages) == 3
+    assert response.pages[0].index == 0
+    assert response.pages[0].markdown == "Page 1 block A\n\nPage 1 block B"
+    assert getattr(response.pages[0], "blocks")[0]["bbox"]["page"] == 1
+    assert response.pages[1].markdown == "Page 2 block A"
+    assert response.pages[2].markdown == "Page 3 block A"
+    assert response._hidden_params["reducto_raw"]["usage"]["credits"] == 3
+
+
+@pytest.mark.asyncio
+async def test_parse_v3_reducto_id_passthrough_skips_upload(
+    disable_aiohttp_transport, respx_mock
+):
+    upload_route = respx_mock.post("https://platform.reducto.ai/upload").respond(
+        json={"file_id": "reducto://should-not-upload.pdf"}
+    )
+    parse_route = respx_mock.post("https://platform.reducto.ai/parse").respond(
+        json=_reducto_parse_response()
+    )
+
+    response = await litellm.aocr(
+        model="reducto/parse-v3",
+        document={
+            "type": "document_url",
+            "document_url": "reducto://already-uploaded.pdf",
+        },
+        api_key="test-key",
+        api_base="https://platform.reducto.ai",
+        retrieval={"chunk_mode": "section"},
+    )
+
+    assert not upload_route.called
+    assert parse_route.called
+    parse_request_body = json.loads(parse_route.calls[0].request.read())
+    assert parse_request_body["input"] == "reducto://already-uploaded.pdf"
+    assert parse_request_body["retrieval"]["chunk_mode"] == "section"
+    assert response.pages[0].markdown.startswith("Page 1 block A")

tests/test_litellm/llms/reducto/test_upload.py

@@ -0,0 +1,213 @@
+import json
+import os
+from unittest.mock import AsyncMock, Mock
+
+import httpx
+import litellm
+import pytest
+
+from litellm.llms.reducto.common import (
+    extract_file_id_or_bytes,
+    upload_bytes_async,
+    upload_bytes_sync,
+)
+
+
+@pytest.fixture()
+def disable_aiohttp_transport(monkeypatch):
+    original_disable_aiohttp = litellm.disable_aiohttp_transport
+    litellm.disable_aiohttp_transport = True
+    litellm.in_memory_llm_clients_cache.flush_cache()
+    monkeypatch.setenv("REDUCTO_API_KEY", "env-reducto-key")
+    try:
+        yield
+    finally:
+        litellm.disable_aiohttp_transport = original_disable_aiohttp
+        litellm.in_memory_llm_clients_cache.flush_cache()
+        os.environ.pop("REDUCTO_API_KEY", None)
+
+
+@pytest.mark.asyncio
+async def test_parse_v3_rejects_plain_http_urls(disable_aiohttp_transport):
+    with pytest.raises(litellm.BadRequestError, match="upload the file first"):
+        await litellm.aocr(
+            model="reducto/parse-v3",
+            document={
+                "type": "document_url",
+                "document_url": "https://example.com/document.pdf",
+            },
+            api_key="test-key",
+            api_base="https://platform.reducto.ai",
+        )
+
+
+@pytest.mark.asyncio
+async def test_parse_v3_image_data_uri_upload_uses_image_mime(
+    disable_aiohttp_transport, respx_mock
+):
+    upload_route = respx_mock.post("https://custom.reducto.test/upload").respond(
+        json={"file_id": "reducto://uploaded-image.png"}
+    )
+    parse_route = respx_mock.post("https://custom.reducto.test/parse").respond(
+        json={
+            "usage": {"num_pages": 1, "credits": 1},
+            "result": {
+                "chunks": [
+                    {
+                        "content": "Image OCR",
+                        "blocks": [{"content": "Image OCR", "bbox": {"page": 1}}],
+                    }
+                ]
+            },
+        }
+    )
+
+    response = await litellm.aocr(
+        model="reducto/parse-v3",
+        document={
+            "type": "file",
+            "file": b"\x89PNG\r\n\x1a\npng",
+            "mime_type": "image/png",
+        },
+        api_key="programmatic-key",
+        api_base="https://custom.reducto.test/",
+    )
+
+    assert upload_route.called
+    assert parse_route.called
+    upload_request = upload_route.calls[0].request
+    assert upload_request.headers["authorization"] == "Bearer programmatic-key"
+    assert b"image/png" in upload_request.read()
+
+    parse_request_body = json.loads(parse_route.calls[0].request.read())
+    assert parse_request_body["input"] == "reducto://uploaded-image.png"
+    assert response.pages[0].markdown == "Image OCR"
+
+
+@pytest.mark.asyncio
+async def test_parse_v3_uses_programmatic_api_key_over_env(
+    disable_aiohttp_transport, respx_mock
+):
+    upload_route = respx_mock.post("https://platform.reducto.ai/upload").respond(
+        json={"file_id": "reducto://uploaded.pdf"}
+    )
+    parse_route = respx_mock.post("https://platform.reducto.ai/parse").respond(
+        json={
+            "usage": {"num_pages": 1, "credits": 1},
+            "result": {
+                "chunks": [
+                    {
+                        "content": "Programmatic auth",
+                        "blocks": [
+                            {"content": "Programmatic auth", "bbox": {"page": 1}}
+                        ],
+                    }
+                ]
+            },
+        }
+    )
+
+    await litellm.aocr(
+        model="reducto/parse-v3",
+        document={
+            "type": "file",
+            "file": b"%PDF-1.4 auth",
+            "mime_type": "application/pdf",
+        },
+        api_key="passed-key",
+        api_base="https://platform.reducto.ai",
+    )
+
+    assert upload_route.calls[0].request.headers["authorization"] == "Bearer passed-key"
+    assert parse_route.calls[0].request.headers["authorization"] == "Bearer passed-key"
+
+
+def test_upload_bytes_sync_uses_shared_client(monkeypatch):
+    captured = {}
+
+    def fake_post(*, url, headers, files, timeout):
+        captured["url"] = url
+        captured["headers"] = headers
+        captured["files"] = files
+        captured["timeout"] = timeout
+        return httpx.Response(
+            200,
+            json={"file_id": "reducto://sync-upload"},
+            request=httpx.Request("POST", url),
+        )
+
+    sync_post = Mock(side_effect=fake_post)
+    monkeypatch.setattr(litellm.module_level_client, "post", sync_post)
+
+    class ForbiddenSyncClient:
+        def __init__(self, *args, **kwargs):
+            raise AssertionError("should not construct")
+
+    monkeypatch.setattr(httpx, "Client", ForbiddenSyncClient)
+
+    file_id = upload_bytes_sync(
+        raw_bytes=b"%PDF-1.4 sync",
+        mime="application/pdf",
+        api_key="sync-key",
+        api_base="https://sync.reducto.test/",
+    )
+
+    assert file_id == "reducto://sync-upload"
+    sync_post.assert_called_once()
+    assert captured["url"] == "https://sync.reducto.test/upload"
+    assert captured["headers"] == {"Authorization": "Bearer sync-key"}
+    assert captured["files"]["file"] == (
+        "document",
+        b"%PDF-1.4 sync",
+        "application/pdf",
+    )
+
+
+@pytest.mark.asyncio
+async def test_upload_bytes_async_uses_shared_aclient(monkeypatch):
+    captured = {}
+
+    async def fake_post(*, url, headers, files, timeout):
+        captured["url"] = url
+        captured["headers"] = headers
+        captured["files"] = files
+        captured["timeout"] = timeout
+        return httpx.Response(
+            200,
+            json={"file_id": "reducto://async-upload"},
+            request=httpx.Request("POST", url),
+        )
+
+    async_post = AsyncMock(side_effect=fake_post)
+    monkeypatch.setattr(litellm.module_level_aclient, "post", async_post)
+
+    class ForbiddenAsyncClient:
+        def __init__(self, *args, **kwargs):
+            raise AssertionError("should not construct")
+
+    monkeypatch.setattr(httpx, "AsyncClient", ForbiddenAsyncClient)
+
+    file_id = await upload_bytes_async(
+        raw_bytes=b"%PDF-1.4 async",
+        mime="application/pdf",
+        api_key="async-key",
+        api_base="https://async.reducto.test/",
+    )
+
+    assert file_id == "reducto://async-upload"
+    async_post.assert_awaited_once()
+    assert captured["url"] == "https://async.reducto.test/upload"
+    assert captured["headers"] == {"Authorization": "Bearer async-key"}
+    assert captured["files"]["file"] == (
+        "document",
+        b"%PDF-1.4 async",
+        "application/pdf",
+    )
+
+
+def test_extract_file_id_or_bytes_raises_on_malformed_data_uri():
+    with pytest.raises(litellm.BadRequestError, match="Invalid Reducto data URI"):
+        extract_file_id_or_bytes("data:application/pdf", model="reducto/parse-v3")
+
+    with pytest.raises(litellm.BadRequestError, match="Invalid Reducto base64 payload"):
+        extract_file_id_or_bytes("data:;base64,!!!not-base64", model="reducto/parse-v3")

tests/test_litellm/test_utils.py

@@ -675,6 +675,7 @@ def test_aaamodel_prices_and_context_window_json_is_valid():
                 "input_dbu_cost_per_token": {"type": "number"},
                 "annotation_cost_per_page": {"type": "number"},
                 "ocr_cost_per_page": {"type": "number"},
+                "ocr_cost_per_credit": {"type": "number"},
                 "code_interpreter_cost_per_session": {"type": "number"},
                 "inference_geo": {"type": "string"},
                 "litellm_provider": {"type": "string"},