fix: Handle missing token details in vLLM/OpenAI-compatible APIs (#2669)

Co-authored-by: claude[bot] <209825114+claude[bot]@users.noreply.github.com>
Co-authored-by: Douwe Maan <[email protected]>

Author: DouweM

pydantic_ai_slim/pydantic_ai/models/openai.py

@@ -1375,11 +1375,21 @@ def _map_usage(response: chat.ChatCompletion | ChatCompletionChunk | responses.R
             ).items()
             if isinstance(value, int)
         }
-        details['reasoning_tokens'] = response_usage.output_tokens_details.reasoning_tokens
+        # Handle vLLM compatibility - some providers don't include token details
+        if getattr(response_usage, 'input_tokens_details', None) is not None:
+            cache_read_tokens = response_usage.input_tokens_details.cached_tokens
+        else:
+            cache_read_tokens = 0
+
+        if getattr(response_usage, 'output_tokens_details', None) is not None:
+            details['reasoning_tokens'] = response_usage.output_tokens_details.reasoning_tokens
+        else:
+            details['reasoning_tokens'] = 0
+
         return usage.RequestUsage(
             input_tokens=response_usage.input_tokens,
             output_tokens=response_usage.output_tokens,
-            cache_read_tokens=response_usage.input_tokens_details.cached_tokens,
+            cache_read_tokens=cache_read_tokens,
             details=details,
         )
     else:

tests/models/cassettes/test_openai_responses/test_openai_responses_usage_without_tokens_details.yaml

@@ -0,0 +1,96 @@
+interactions:
+- request:
+    headers:
+      accept:
+      - application/json
+      accept-encoding:
+      - gzip, deflate
+      connection:
+      - keep-alive
+      content-length:
+      - '84'
+      content-type:
+      - application/json
+      host:
+      - api.openai.com
+    method: POST
+    parsed_body:
+      input:
+      - content: What is 2+2?
+        role: user
+      model: gpt-4o
+      stream: false
+    uri: https://api.openai.com/v1/responses
+  response:
+    headers:
+      alt-svc:
+      - h3=":443"; ma=86400
+      connection:
+      - keep-alive
+      content-length:
+      - '1369'
+      content-type:
+      - application/json
+      openai-organization:
+      - pydantic-28gund
+      openai-processing-ms:
+      - '899'
+      openai-project:
+      - proj_dKobscVY9YJxeEaDJen54e3d
+      openai-version:
+      - '2020-10-01'
+      strict-transport-security:
+      - max-age=31536000; includeSubDomains; preload
+      transfer-encoding:
+      - chunked
+    parsed_body:
+      background: false
+      created_at: 1756246493
+      error: null
+      id: resp_68ae31dd0edc819da9e77ad3cdb128770306cb62070aed80
+      incomplete_details: null
+      instructions: null
+      max_output_tokens: null
+      max_tool_calls: null
+      metadata: {}
+      model: gpt-4o-2024-08-06
+      object: response
+      output:
+      - content:
+        - annotations: []
+          logprobs: []
+          text: 2 + 2 equals 4.
+          type: output_text
+        id: msg_68ae31ddcef0819da13eb00d2b393eb60306cb62070aed80
+        role: assistant
+        status: completed
+        type: message
+      parallel_tool_calls: true
+      previous_response_id: null
+      prompt_cache_key: null
+      reasoning:
+        effort: null
+        summary: null
+      safety_identifier: null
+      service_tier: default
+      status: completed
+      store: true
+      temperature: 1.0
+      text:
+        format:
+          type: text
+        verbosity: medium
+      tool_choice: auto
+      tools: []
+      top_logprobs: 0
+      top_p: 1.0
+      truncation: disabled
+      usage:
+        input_tokens: 14
+        output_tokens: 9
+        total_tokens: 23
+      user: null
+    status:
+      code: 200
+      message: OK
+version: 1

tests/models/test_openai_responses.py

@@ -29,7 +29,7 @@
 from pydantic_ai.output import NativeOutput, PromptedOutput, TextOutput, ToolOutput
 from pydantic_ai.profiles.openai import openai_model_profile
 from pydantic_ai.tools import ToolDefinition
-from pydantic_ai.usage import RequestUsage
+from pydantic_ai.usage import RequestUsage, RunUsage
 
 from ..conftest import IsDatetime, IsStr, TestEnv, try_import
 from ..parts_from_messages import part_types_from_messages
@@ -1076,3 +1076,16 @@ async def test_openai_responses_verbosity(allow_model_requests: None, openai_api
     agent = Agent(model=model, model_settings=OpenAIResponsesModelSettings(openai_text_verbosity='low'))
     result = await agent.run('What is 2+2?')
     assert result.output == snapshot('4')
+
+
+async def test_openai_responses_usage_without_tokens_details(allow_model_requests: None, openai_api_key: str):
+    # The VCR cassette was manually modified to remove the input_tokens_details and output_tokens_details fields.
+    provider = OpenAIProvider(api_key=openai_api_key)
+    model = OpenAIResponsesModel('gpt-4o', provider=provider)
+
+    agent = Agent(model=model)
+    result = await agent.run('What is 2+2?')
+
+    assert result.usage() == snapshot(
+        RunUsage(input_tokens=14, output_tokens=9, details={'reasoning_tokens': 0}, requests=1)
+    )