GPT-OSS 2, served as an API
The successor to the models that brought OpenAI back to open weights, and the most-called family in this catalog. Point the OpenAI SDK at our base URL and call it per token, no contract and no minimums.
- Apache 2.0 expected
- Reasoning and tool calling
- Structured output
- Per token, no contract
jk. It is not out yet.
OpenAI has not announced a second GPT-OSS generation. One aggregator describes one, but it contradicts OpenAI's own documentation, and we are not going to publish a spec sheet off that. Both current GPT-OSS models are live here now.
Below is every rumor we could source, what OpenAI has actually confirmed, and the model you can call in the meantime. When the real one ships, this page becomes its page at this URL.
The GPT-OSS 2 rumor mill.
Each row labelled by how much it is worth. Last checked September 24, 2026.
- Release status
- ConfirmedUnreleased. OpenAI's open-weight line is gpt-oss-120b and gpt-oss-20b under Apache 2.0: mixture-of-experts at 117B and 21B total parameters, activating 5.1B and 3.6B per token, with context up to 128K. OpenAI, gpt-oss model card
- Leaks
- No signalNo official OpenAI statement about a successor, and nothing we would call a credible leak.
- The claim we do not believe
- UnverifiedOne aggregator describes a June 9, 2026 refresh with three sizes (gpt-oss-8b, gpt-oss-20b-a3b, gpt-oss-120b-a12b) at 256K context under an “OpenAI Open Weight License 1.0”. It contradicts OpenAI's own documentation, which still lists two models under Apache 2.0. Repeating it here because it is circulating, not because it checks out.
- The name
- UnverifiedOurs, and a weaker guess than most on this site. OpenAI shipped the first line with no generation number at all, so a successor could be called nearly anything. This page follows OpenAI's next open-weight model whatever it is named.
- The pressure a successor would answer
- UnverifiedBy mid-2026 the open-weight leaderboard is led by much larger Chinese-lab models. That is the competitive case for a refresh, and it is analysis rather than reporting.
- Scope
- ConfirmedOpenAI's open-weight line only. GPT-5 and the hosted models are not licensed for third-party serving, we do not host them, and no page here will claim otherwise.
- How it would serve here
- ConfirmedSelf-hosted on capacity we operate, the same path both current GPT-OSS models run on today.
- Price
- No signalUnknown until we have the model and can measure what it costs to serve.
What you can call right now.
The OpenAI models live on the API today, at the rates they bill at.
Full inference catalogGPT-OSS 120B
The larger open-weight model, with reasoning, tool calling, and structured output. The most popular model in the catalog.
Input / output per 1M tokens
$0.15 / $0.60
GPT-OSS 20B
The smaller variant at lower rates, for high-throughput, latency-sensitive work that still needs tool calling.
Input / output per 1M tokens
$0.05 / $0.20
Why it shows up here early.
Not a promise about a date. A description of the path a new open model takes to get behind this API.
GPT-OSS is the most-called family here
Both sizes run on capacity we operate and carry the most traffic in the catalog, so a successor is the release we are most prepared to stand up.
Open weights make this a days problem
The work between a public weights drop and a served endpoint is engine support and capacity, and both are standing for this family.
Nothing to migrate when it arrives
The SDK is already OpenAI's. Moving to the next open-weight model means changing one string, with no client rewrite and no new vendor.
One email, the day it is callable.
Leave an address and we send exactly one message when GPT-OSS 2 is serving here, with the model id and the rate. Nothing else goes to this list.
GPT-OSS 2, answered.
Is GPT-OSS 2 out?
No. As of September 2026 OpenAI has not announced a successor to gpt-oss-120b and gpt-oss-20b. Both of those are live on OpenRelay today.
I saw a post about new gpt-oss sizes. Is that real?
We could not verify it. One aggregator describes a June 2026 refresh with three sizes and a new licence, and it contradicts OpenAI's own documentation, which still lists two models under Apache 2.0. Until OpenAI or a Hugging Face model card says otherwise, treat it as unconfirmed.
When will OpenAI release another open-weight model?
Unannounced. OpenAI published the first GPT-OSS weights to Hugging Face; if a successor lands the same way, this page becomes the live model page and the notify list gets an email.
Will OpenRelay serve it?
If it ships with open weights, yes, and fast. GPT-OSS is the most-used family in this catalog, so the serving path, the capacity, and the demand are all already here.
Can I get GPT-5 through this API?
No. GPT-5 and OpenAI's other hosted models are not licensed for third-party serving, so no provider outside OpenAI can offer them. This page is about the open-weight line, which is the part anyone can host.
What should I use until then?
GPT-OSS 120B for agents, tool use, and harder reasoning, or GPT-OSS 20B when throughput and latency matter more. Both are on the cards above at their live rates.
More models
The account outlasts the model.
One API key, one base URL, per-token billing. Whatever OpenAI ships next, adopting it is a one-line change rather than a new vendor.