Can the OpenAI Agents SDK use Claude or Gemini?

Yes. Set an AsyncOpenAI client pointed at https://api.cognocient.com/v1 as the default with set_default_openai_client, switch to the Chat Completions API with set_default_openai_api('chat_completions'), and give agents any gateway model name such as claude-sonnet-4-6.

The SDK uploads traces to OpenAI using the default client's key. With a Cognocient AI key that upload fails, so either disable tracing or give the tracer a real OpenAI key with set_tracing_export_api_key.

Do hosted tools like WebSearchTool work?

No. Hosted tools run on OpenAI's servers and need OpenAI's stateful Responses API. Function tools, handoffs, guardrails and structured outputs work.

OpenAI Agents SDK with Claude, Gemini and other models: cost tracking and run budgets through a gateway

Run OpenAI Agents SDK agents on any model by pointing the default client at an OpenAI-compatible gateway. Track cost per agent and per run, and stop runaway agents with a dollar budget.

The Agents SDK is built on the OpenAI client, so the gateway plugs in as the default client. Every agent, handoff and tool-calling turn then goes through it, on any model.

Setup

# pip install openai-agents
from openai import AsyncOpenAI
from agents import (Agent, Runner, function_tool, set_default_openai_api,
                    set_default_openai_client, set_tracing_disabled)
 
client = AsyncOpenAI(
    api_key="sk-cog-YOUR-KEY",
    base_url="https://api.cognocient.com/v1",
    default_headers={"X-Cost-Feature": "support-triage"},
)
set_default_openai_client(client, use_for_tracing=False)
set_default_openai_api("chat_completions")   # portable across every provider
set_tracing_disabled(True)                     # or set_tracing_export_api_key("sk-...openai...")
 
@function_tool
def lookup_order(order_id: str) -> str:
    return f"Order {order_id}: shipped"
 
billing = Agent(name="Billing", instructions="Answer billing questions.", model="gpt-4o-mini")
triage = Agent(
    name="Triage",
    instructions="Route the user to the right specialist, or look up their order.",
    model="claude-sonnet-4-6",
    tools=[lookup_order],
    handoffs=[billing],
)
 
result = Runner.run_sync(triage, "Where is order 1042?")
print(result.final_output)

Each agent can use a different model and provider; the gateway records each call against its real model and price.

Cost per run

Tag a run by creating the client per run, or by passing headers through ModelSettings:

from agents import ModelSettings
 
triage = Agent(
    name="Triage",
    model="claude-sonnet-4-6",
    model_settings=ModelSettings(extra_headers={"X-Cost-Run-ID": run_id, "X-Cost-Run-Budget": "0.75"}),
)

With X-Cost-Run-Budget the gateway stops the run with a 429 (run_budget_exceeded) once it has spent that amount. The SDK raises it as an exception from Runner.run, which you can catch and turn into a fallback answer. max_turns limits turns; the run budget limits dollars.

Responses API mode

You can keep the SDK on the Responses API with set_default_openai_api("responses"). The gateway's /v1/responses is stateless: the SDK's default behaviour of sending the full input each turn works, but previous_response_id and hosted tools don't. Chat Completions mode is the safer default for non-OpenAI models.

OpenAI Agents SDK with Claude, Gemini and other models: cost tracking and run budgets through a gateway

Setup

Cost per run

On this page