OpenAI

Models and offerings #

Who they are

This hub is for engineers calling a hosted model, whether or not they end up calling OpenAI’s, because the request shape OpenAI chose in 2023 is now the one almost everyone speaks. OpenAI was founded in December 2015 as a non-profit research lab, published the GPT papers from 2018, and launched ChatGPT in November 2022, which started the current wave. In 2025 it reorganised so that a public benefit corporation, OpenAI Group PBC, operates the business under the control of the non-profit OpenAI Foundation.

In the field it is the reference point: the largest consumer product, the API other providers describe themselves as compatible with, and the maker of the gpt-oss models that put a credible open-weight reasoning model on Apache-2.0 terms. It is also the most closed of the big three labs about model sizes. No GPT model since GPT-3 has a disclosed parameter count.

Tooling and APIs

pip install openai installs the official Python SDK, with Node, .NET, Java and Go equivalents. The Responses API is the current primary endpoint and the one new features land on. Chat Completions remains supported and is what most third-party code targets. Keys come from the platform dashboard and go in OPENAI_API_KEY. The SDK works from any OS, but the models run only in OpenAI’s cloud and through Microsoft Foundry.

The minimal call:

# needs-api
import os
from openai import OpenAI

client = OpenAI(api_key=os.environ["OPENAI_API_KEY"])
response = client.responses.create(
    model="gpt-6.1-terra",
    input="In one sentence, what does a learning rate control?",
)
print(response.output_text)

Models and prices at the time of writing, per million tokens, from OpenAI’s pricing page: GPT-6 Astra 10 USD in and 50 USD out, GPT-6.1 Sol 2 and 10, GPT-6.1 Terra 1 and 5, GPT-6 Luna 0.10 and 0.50. Cached input is a tenth of the base price on the GPT-6 generation, and the Batch API halves everything for asynchronous jobs. All four have a 1.05 M token context window and 128 k output. The previous GPT-5.x models are still served at their old prices, and o3 and GPT-4o remain for compatibility. Embeddings cost 0.02 USD per million tokens for text-embedding-3-small. Fine-tuning is closed to new users according to the same page. Release dates for the GPT-6 models are not stated on the model pages, and parameter counts are undisclosed.

The open-weight models are a different proposition. gpt-oss-120b (117 B total, 5.1 B active, 128 k context, Apache-2.0, 5 August 2025) fits one 80 GB GPU in its native MXFP4 format, and gpt-oss-20b (21 B total, 3.6 B active) runs on a 16 GB laptop. Both are on Hugging Face and in Ollama.

Pain points: pricing tiers multiply (standard, batch, flex, fast, priority), model names change every few months, and the Responses API’s stateful conversation objects are convenient until you need to reproduce a request exactly.

Migration notes

To another provider: Anthropic has its own Messages format with a top-level system field and content blocks, and its SDK is a different package. Google, Mistral, Meta, NVIDIA NIM, Microsoft Foundry Local, vLLM, llama.cpp and Ollama all expose an OpenAI-compatible endpoint, so the usual route is to keep the openai SDK and change base_url and model. Tool-calling schemas, JSON-mode flags and reasoning-effort parameters are where compatibility breaks, so test those first. LiteLLM exists to paper over the remaining differences.

Structured output: the response_format and text.format JSON schema options are the most reliable way to get parseable output, and our prompt patterns reference shows the equivalent on each provider.

From gpt-oss to a hosted GPT or back: the open models use a “harmony” chat format with explicit channels for reasoning and final answers. Serving stacks handle it, but if you template prompts by hand you must match it.

Watch for

Lock-in is mostly in stateful features: stored conversations, the file and vector store APIs, fine-tuned models and assistants exist only on OpenAI’s side. Keep your own copy of prompts and transcripts.

Rate limits are per model and per tier, and a new account starts low, with limits that rise with spend. Deprecations are announced with a few months’ notice and old model ids stop working, so pin dated snapshots and read the deprecation page. API data is not used for training by default for business customers, but ChatGPT consumer conversations can be unless you opt out, and zero-data-retention requires a contract. The usage policies restrict some domains (the “Trusted Access” models on the pricing page are gated for that reason).

Benchmark numbers in launch posts are the vendor’s own and should be read as such. Confident wrong output (hallucination) is not solved by any model on this page. The gpt-oss weights are open but the training data and code are not, so they are open weights, not open source.