GPT family

OpenAI’s GPT text models, including the GPT-6, GPT-5.x, GPT-4.1, GPT-4o and Codex series. They are called through the OpenAI-compatible API, which supports both Chat Completions and Responses; the official OpenAI SDK works once you change base_url.

Common tasks

Enable reasoning reasoning_effort values, reasoning tokens, reasoning summaries.Web search The web_search tool on the Responses API.Prompt caching Automatic caching and confirming hits.Structured output & PDF json_schema output and PDF file input.Troubleshooting errors Context exceeded, max_tokens too large, empty responses, unsupported endpoint.Use in Codex Setup and common questions.

Models

All models support tool calling and structured output; all except GPT-5 nano accept image input. For the full model list and live prices, see the model page , or query the Models API.

Common model IDs (copy them exactly)

GPT-6 Solopenai/gpt-6-sol

GPT-5.4openai/gpt-5.4

GPT-5.4 miniopenai/gpt-5.4-mini

GPT-4.1 miniopenai/gpt-4.1-miniNo reasoning; suited to simple tasks

You can omit the openai/ prefix, e.g. gpt-5.4. The model field in responses always includes the prefix.

Choosing an endpoint

Key pointThe Codex series and GPT-5.4 Pro support only the Responses API; all other models work with both.

	Chat Completions	Responses
Endpoint	`POST https://api.ofox.io/v1/chat/completions`	`POST https://api.ofox.io/v1/responses`
Input	`messages`	`input` (a string or an array of messages)
Output limit	`max_completion_tokens` (`max_tokens` also accepted)	`max_output_tokens`
Reasoning effort	`reasoning_effort`	`reasoning.effort`
Reasoning summary	Not returned	`reasoning.summary` — see Reasoning summaries
Web search	Not supported	`web_search` tool
Models available only here	—	`openai/gpt-5.4-pro`, `openai/gpt-5.3-codex`, `openai/gpt-5.2-codex`

For the full request parameters, see the Chat Completions API and the Responses API.

Providers

Provider	Models
Azure AI Foundry (`azure_foundry`)	All models; default provider
OpenAI (`openai`)	GPT-6 Sol, GPT-6 Luna, GPT-6 Astra, GPT-5.6 Sol, GPT-5.6 Terra, GPT-5.6 Luna

In our tests, both providers behaved the same for reasoning, caching and web search. To select a provider, add the X-OfoxAI-Provider-Type: openai request header; for other methods, see Provider routing.

Examples

Python (Chat Completions)

gpt_chat.py


import os
from openai import OpenAI
 
client = OpenAI(base_url="https://api.ofox.io/v1", api_key=os.environ["OFOX_API_KEY"])
 
resp = client.chat.completions.create(
    model="openai/gpt-5.4",
    messages=[
        {"role": "system", "content": "You are a senior backend engineer. Keep answers concise."},
        {"role": "user", "content": "Explain idempotency in one sentence."},
    ],
)
print(resp.choices[0].message.content)

Python (Responses)

gpt_responses.py


import os
from openai import OpenAI
 
client = OpenAI(base_url="https://api.ofox.io/v1", api_key=os.environ["OFOX_API_KEY"])
 
resp = client.responses.create(
    model="openai/gpt-5.4",
    instructions="You are a senior backend engineer. Keep answers concise.",
    input="Explain idempotency in one sentence.",
)
print(resp.output_text)

cURL

Terminal


curl https://api.ofox.io/v1/chat/completions \
  -H "Authorization: Bearer $OFOX_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "openai/gpt-5.4",
    "messages": [{"role": "user", "content": "Explain idempotency in one sentence."}]
  }'

Prompt caching

Key pointThe GPT family caches automatically with no changes to your request. Put fixed content at the start of the request to improve the hit rate.

When a request’s prefix matches a recent request and is long enough, it hits the cache automatically, and the cached portion is billed at the cache read rate.

Item	Details
Confirming hits	In Chat Completions, `usage.prompt_tokens_details.cached_tokens` is greater than 0; in Responses, `usage.input_tokens_details.cached_tokens`
Writes	The GPT-6 and GPT-5.6 series return `prompt_tokens_details.cache_write_tokens` on the first request, billed at the cache write rate
Minimum length	Prefixes that are too short are not cached; for the minimum length, see the OpenAI docs
Reliability	Caching is best-effort; requests with the same prefix may still miss

For general information on caching, see Prompt caching.

Web search

Add web_search to tools on the Responses API, and the model runs searches on its own and cites sources. Chat Completions does not support web search.

gpt_web_search.py


import os
from openai import OpenAI
 
client = OpenAI(base_url="https://api.ofox.io/v1", api_key=os.environ["OFOX_API_KEY"])
 
resp = client.responses.create(
    model="openai/gpt-5.4",
    tools=[{"type": "web_search"}],
    input="What is the latest stable version of Python?",
)
print(resp.output_text)
print("Searches:", sum(1 for item in resp.output if item.type == "web_search_call"))

Each web_search_call item in output is one search, and the message item is the final answer. Searches are billed per use; see the model page for the rate.

Common errors

Key pointParameter errors return 400 with a reason. Long reasoning occasionally causes 502 or 504; use streaming and set a long timeout.

HTTP · type	Error message	Cause	Fix
400 `invalid_request_error`	`This model's maximum context length is 128000 tokens. However, your messages resulted in 152008 tokens.`	Input exceeds the model’s context window	Shorten the input, or switch to a model with a longer context
400 `invalid_request_error`	`max_tokens is too large: 100000. This model supports at most 16384 completion tokens`	Output limit exceeds the model’s maximum	Lower `max_completion_tokens`
400 `invalid_request_error`	`Invalid value: 'bogus'. Supported values are: 'low', 'medium', 'high', and 'xhigh'.`	The `reasoning_effort` value is not supported by this model	Use one of the values listed in the error message
400 `endpoint_not_supported`	`Model 'openai/gpt-5.4-pro' does not support the /v1/chat/completions endpoint on this platform.`	The model supports only the Responses API	Use `/v1/responses`
404 `model_not_found`	`Model 'openai/gpt-9' not found`	Model ID misspelled, or the model has been retired	Copy the model ID from this page
429 `rate_limit_error`	`Rate limit reached for the requested model. Please retry after a short delay.`	Too many requests	Retry later — see Rate limits
502 `upstream_error`	`The model service is temporarily unavailable. Please retry or switch to another model.`	Upstream service temporarily unavailable	Retry later, or switch to another model
504 `gateway_timeout`	`The upstream service did not respond in time.`	Reasoning took too long and upstream did not respond in time	Use streaming, or lower the reasoning effort
200, `finish_reason: "length"` with an empty answer	—	Reasoning used up `max_completion_tokens`	Increase `max_completion_tokens`, or lower the reasoning effort

For other general errors, see Error Handling.

Official docs

The official docs describe behavior when calling OpenAI directly. When calling through OfoxAI, model IDs and endpoints on this page take precedence.

FAQ

How do I write a GPT model ID?

The standard format is openai/gpt-5.4. You can also omit the openai/ prefix and write gpt-5.4.

Should I use Chat Completions or Responses?

Either works. The Codex series and GPT-5.4 Pro support only Responses, and you also need Responses for reasoning summaries or web search. Projects with existing Chat Completions code can keep using it.

How do I enable reasoning for GPT?

In Chat Completions, pass reasoning_effort (e.g. "high"); in Responses, pass reasoning: {"effort": "high"}. Some models do not reason by default, so pass it explicitly when you need it.

Why is the GPT response empty with finish_reason length?

Reasoning tokens count toward max_completion_tokens, and once reasoning uses up the budget no answer is produced. Increase max_completion_tokens or lower reasoning_effort.

What does the error does not support the /v1/chat/completions endpoint mean?

The model supports only the Responses API, e.g. gpt-5.4-pro, gpt-5.3-codex and gpt-5.2-codex. Use POST /v1/responses instead.

Can I see the GPT reasoning process?

Use the Responses API and pass reasoning: {"effort": "medium", "summary": "auto"}; the reasoning items in output will contain a reasoning summary. Chat Completions does not return reasoning content.

Can GPT search the web?

Yes. Use the Responses API and add {"type": "web_search"} to tools. Chat Completions does not support web search.

Do I need to enable GPT prompt caching manually?

No. When the prefix is long enough and matches a recent request, it hits the cache automatically; the cached amount is in cached_tokens in usage.

What should I do if GPT requests time out or return 504?

High-effort reasoning can take a long time. Use streaming (stream: true), set the client timeout to 600 seconds or more, or lower reasoning_effort.