OpenAI / lineage

OpenAI gpt-oss family

OpenAI’s gpt-oss checkpoints are separate downloadable models. The weights, serving runtime and hosted endpoint each bring their own requirements.

Family research reviewed · Official source notes below

Standard models
3
Independently rated
2 / 3
Known token rates
3 / 3
Global balanced frontier
1

This collection organises accepted catalogue identities. It does not certify equal weights, a shared licence, benchmark equivalence or current availability on every serving platform.

Compare exact catalogue rows

USD per 1M tokens. In / out are listed base rates; expand billing conditions before planning a request. LMArena scores measure the general preference board, not success on your engineering task. Unknown stays unknown.

Catalogue members

ModelLMArenaInput / output $/MContext / max outputCapabilities & conditions
gpt-oss-safeguard-20bopenai/gpt-oss-safeguard-20bUnrated$0.075 / $0.300Listed base rates
Billing details
Cached input
$0.037
Separate reasoning
Not separately listed

Tool, media, cache-write and service-tier charges may add dimensions. Read the full model evidence.

131,072Output 65,536textTools yes · reasoning yesWeights unknownReleased 2025-10-29
gpt-oss-120bopenai/gpt-oss-120b1365.9±4.4 interval half-widthEffort: default$0.150 / $0.600Listed base rates
Billing details
Cached input
$0.075
Separate reasoning
Not separately listed

Tool, media, cache-write and service-tier charges may add dimensions. Read the full model evidence.

131,072Output 65,536textTools yes · reasoning yesOpen weightsReleased 2025-08-05
gpt-oss-20bopenai/gpt-oss-20b1287.3±6.4 interval half-widthEffort: default$0.018 / $0.090Listed base rates
Billing details
Cached input
Unknown
Separate reasoning
Not separately listed

Tool, media, cache-write and service-tier charges may add dimensions. Read the full model evidence.

131,072Output 32,768textTools yes · reasoning yesWeights unknownReleased 2025-08-05

Free and batch variants

These rows have different access or processing conditions. They are excluded from the standard-row counts above. A variant without a standard base remains a separately named variant.

Sources and scope

  • OpenAI model catalogue ↗

    Exact named models on the vendor platform; not a promise about every reseller

    Checked 2026-09-22 · official document opened and inspected

Source review does not refresh a price or transplant a benchmark score. Read the method and correction policy.

Questions this page answers

Are the OpenAI gpt-oss family models interchangeable?

No. These are exact catalogue rows with separate capability, pricing and rating evidence. Compare the checkpoint, serving endpoint, reasoning settings and required input types. A family name does not establish equivalent behaviour.

How many OpenAI rows are compared here?

3 standard rows, of which 2 have an accepted LMArena general score and 3 have known input and output token rates. 1 related free or batch variants are listed separately.

What do the prices and frontier counts mean?

Prices below are listed US dollars per million input or output tokens, before request-specific context tiers and other billing dimensions. The frontier count comes from the global balanced view. Neither a family price range nor a benchmark difference proves your task will be cheaper or more successful.

How current are the official notes?

The linked family research was reviewed 2026-09-22. That date is separate from each accepted catalogue observation and each vendor’s effective date. An official note applies to its named platform, not automatically to every reseller.

Evidence & Ask