[BUG] OpenAI API call fails with "response_format type is unavailable now" when using Deepseek

[Fiftyw3bs]

Description

Summary

When running a CrewAI flow that makes LLM calls, I encounter the following error:

ERROR:root:OpenAI API call failed: Error code: 400 - {'error': {'message': 'This response_format type is unavailable now', 'type': 'invalid_request_error', 'param': None, 'code': 'invalid_request_error'}}

The error is accompanied by a warning about event pairing mismatch and leads to a complete failure of the LLM call.

Steps to Reproduce

1. Configure CrewAI with Deepseek V4 Pro.
2. Use a crew/flow that relies on structured output or tool calling
3. Run the flow.

Expected behavior

The LLM call should succeed without errors, and the crew should execute normally.

Screenshots/Code snippets

from crewai import Agent, Task, Crew
from crewai.llm import LLM

llm = LLM(
    model="deepseek/deepseek-v4-pro",
    base_url="https://api.deepseek.com",
    api_key=os.getenv("DEEPSEEK_API_KEY"),
)

agent = Agent(role="test", goal="test", backstory="test", llm=llm)
task = Task(description="Say hello", expected_output="A greeting")
crew = Crew(agents=[agent], tasks=[task])
result = crew.kickoff()

Operating System

macOS Sonoma

Python Version

3.12

crewAI Version

1.14.5

crewAI Tools Version

1.14.5

Virtual Environment

Venv

Evidence

ERROR:root:OpenAI API call failed: Error code: 400 - {'error': {'message': 'This response_format type is unavailable now', 'type': 'invalid_request_error', 'param': None, 'code': 'invalid_request_error'}}
ERROR:root:OpenAI API call failed: Error code: 400 - {'error': {'message': 'This response_format type is unavailable now', 'type': 'invalid_request_error', 'param': None, 'code': 'invalid_request_error'}}
[CrewAIEventsBus] Warning: Event pairing mismatch. 'llm_call_failed' closed 'agent_execution_started' (expected 'llm_call_started')
╭─────────────────────────────────────────────────────────────────────── ❌ LLM Error ────────────────────────────────────────────────────────────────────────╮
│  LLM Call Failed                                                                                                                                            │
│  Error: OpenAI API call failed: Error code: 400 - {'error': {'message': 'This response_format type is unavailable now', 'type': 'invalid_request_error',    │
│  'param': None, 'code': 'invalid_request_error'}}                                                                                                           │
╰─────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────╯
An unknown error occurred. Please check the details below.
Error details: Error code: 400 - {'error': {'message': 'This response_format type is unavailable now', 'type': 'invalid_request_error', 'param': None, 'code': 'invalid_request_error'}}
╭─────────────────────────────────────────────────────────────────────── ❌ LLM Error ────────────────────────────────────────────────────────────────────────╮
│  LLM Call Failed                                                                                                                                            │
│  Error: OpenAI API call failed: Error code: 400 - {'error': {'message': 'This response_format type is unavailable now', 'type': 'invalid_request_error',    │
│  'param': None, 'code': 'invalid_request_error'}}                                                                                                           │
╰─────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────╯
ERROR:crewai.flow.flow:Error executing listener call_llm_native_tools: Error code: 400 - {'error': {'message': 'This response_format type is unavailable now', 'type': 'invalid_request_error', 'param': None, 'code': 'invalid_request_error'}}

Possible Solution

None

Additional context

None

[FuturMix]

This is a known issue in CrewAI when using non-OpenAI providers. The root cause: CrewAI's native tool-calling loop sends response_format: {"type": "json_object"} (or structured output params) to the LLM on every call. OpenAI supports this, but most other providers (Deepseek, Gemini, Ollama, vLLM) either reject it or handle it differently.

Immediate workaround — disable structured output in the task:

# Instead of output_pydantic, parse output manually
task = Task(
    description="Say hello",
    expected_output="A greeting in plain text",
    # Don't use output_pydantic or output_json
)

If you need structured output, parse it after the crew finishes:

result = crew.kickoff()
# Parse the text output into your Pydantic model
parsed = YourModel.model_validate_json(result.raw)

Why this happens: Issue #5472 documents this in detail — output_pydantic leaks response_model/response_format into the agent's tool-calling loop, causing non-OpenAI LLMs to fail or skip tools entirely. PR #5821 has a fix that scopes response_model to only the final output step (CI green, awaiting merge).

Alternative approach — route through an OpenAI-compatible proxy that handles response_format translation:
Some API proxies (LiteLLM, FuturMix, etc.) can strip or translate unsupported parameters before forwarding to the upstream provider. This lets you use the same CrewAI code across providers without modifying task configs:

llm = LLM(
    model="deepseek/deepseek-v4-pro",
    base_url="https://your-proxy.example.com/v1",
    api_key=os.getenv("PROXY_API_KEY"),
)

But the cleanest fix is waiting for PR #5821 to merge — it properly separates the output formatting from the tool-calling loop.

---

Disclosure: we run FuturMix, an OpenAI-compatible API with 20+ models.

[norika1207-lab]

One practical way to make this easier to evaluate is to turn the report into a small contract test around the runtime boundary, not only the user-facing symptom.

For Deepseek response_format unavailable when using structured output, I would check three things: first, the failing input is preserved as a minimal fixture; second, the SDK emits the same observable event/result shape before and after the fix except for the intended change; third, the trace or serialized session contains enough metadata to explain what happened after a resume or retry.

The part that matters for agent systems is durability: if a tool call, memory write, handoff, or structured-output step fails, downstream code should not have to infer state from natural-language output. A focused regression test around that boundary would make the issue much easier to close with confidence.