Reasoning Models & reasoning_effort

Reasoning

Reasoning models think before answering. reasoning_effort is the portable control — it is translated into each vendor's own dial, including Anthropic's thinking-token budget — and each vendor's native spelling is also accepted as sent.

Parameter	Type	Required	Description
reasoning_effort	string	Optional	How much the model should think. One of:`minimallowmediumhigh` On Claude this becomes a thinking-token budget.
thinking	object	Optional	Anthropic's native control on Claude models: `{type: "enabled"\|"adaptive"\|"disabled", budget_tokens}`. On the Claude generations that set thinking depth with an effort level, a token budget is converted to one for you.

import os
from openai import OpenAI

client = OpenAI(
    api_key=os.environ["OXYY_API_KEY"],
    base_url="https://api.oxyy.ai/v1"
)

response = client.chat.completions.create(
    model="claude-opus-4-8",
    messages=[{"role": "user", "content": "Prove that sqrt(2) is irrational."}],
    reasoning_effort="high",  # minimal | low | medium | high
)

msg = response.choices[0].message
print(msg.reasoning_content)   # the model's thinking, when it emits any
print(msg.content)
print(response.usage.completion_tokens_details.reasoning_tokens)

# Anthropic-style thinking controls are accepted on Claude models.
# Newer Claude generations get the budget as an effort level.
curl https://api.oxyy.ai/v1/chat/completions \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer $OXYY_API_KEY" \
  -d '{
    "model": "claude-opus-4-8",
    "messages": [{"role": "user", "content": "Plan a migration."}],
    "max_tokens": 8000,
    "thinking": {"type": "enabled", "budget_tokens": 4000}
  }'

Field on the response	What it is
`message.reasoning_content`	The thinking, as text, when the model emits any
`message.reasoning_details`	Structured reasoning blocks. Send them back unmodified in a tool loop so signatures and encrypted reasoning survive.
`usage.completion_tokens_details.reasoning_tokens`	Thinking tokens, counted inside `completion_tokens` and billed at the output rate