See how ChatGPT, Perplexity and Google AI Overviews describe you today

Free AI Visibility Audit
Surge45°
OpenAI model°

gpt-oss-20b

gpt-oss-20b is an open-weight text model from OpenAI, released under the Apache 2.0 licence, with 21 billion total parameters and a Mixture-of-Experts design that activates 3.6 billion of them per forward pass.

OpenAIVerified 24/09/2026Released 05/08/2025

01 / decision

Decision snapshot

Each figure sits next to the middle of the field, so it can be read as dear or cheap, wide or narrow, rather than floating on its own. Compared against the 126 models we track, not against an absolute standard.

Capability

Not measured

No published benchmark scores yet

Input / 1M tokens

$0.02

field median $0.30

Output / 1M tokens

$0.09

field median $1.25

Context

131K

field median 524K

02 / overview

What gpt-oss-20b is, and when to reach for it

Four questions, answered separately, because somebody arrives at one of them rather than at the top of the page.

What it is

The job it was built for, and the job it is not for.

gpt-oss-20b is a small open-weight language model built to be downloaded and run rather than rented, with the Apache 2.0 licence putting no restriction on commercial use or fine-tuning. The sparse Mixture-of-Experts layout means only a fraction of its parameters fire on any given token, so it behaves like a much smaller model at inference time. It handles text only, so anything involving images, audio or document vision needs a different model.

When it arrived, and when to use it

Where it sits in its line, and when a sibling is the better pick.

gpt-oss-20b first appeared in August 2025 as the smaller of OpenAI's open-weight releases. It is the right pick when you need weights you control, on your own hardware or a host of your choosing, and the work is routine: classification, extraction, summarising, drafting, tool calls in a loop. It is the wrong pick when the task needs genuine reasoning depth, long multi-step planning or anything beyond text, where a frontier hosted model still earns its cost.

How you reach it

The API, the apps it powers, and what its limits let you do.

Because the weights are open, gpt-oss-20b is reached however you choose to serve it: locally, on your own GPUs, or through the many inference providers who host it behind an OpenAI-compatible API. The context window is large enough to hold a long document set or a full support thread alongside instructions, and the output ceiling is generous enough for long structured responses rather than short replies. Everything in and out is text.

Why it matters

What changes because this exists, or why it does not.

An open-weight model at this size makes it practical to run high-volume, low-stakes AI work without sending data to a third party or metering every call. Teams that previously batched work to control spend can now run it continuously, and teams with data residency or privacy constraints get a model they can host inside their own boundary. None of this is new as a category, but the licence and the small active parameter count make it unusually easy to justify.

Follows gpt-oss-120b, and is the newest in its line. See the whole line.

03 / evidence

How much of this is verified

Split by category, so a strong number never hides a thin evidence base. Verified means we read it on the benchmark's own published results; a provider's claim about its own model is shown and labelled rather than dropped.

No published benchmark scores for this model yet.

We publish a score only where we can link the result to where it was published. Until a benchmark result for this model exists in a source we read, this section stays empty rather than being filled with a provider's marketing figure.

How we decide what counts as evidence

04 / ledger

Benchmark ledger

Every published row, grouped by category, each compared with the best published score on the same benchmark. 'Is 64% good' is a question nobody can answer; '26 points behind the leader' is one anybody can.

Nothing in the ledger yet.

Each row here carries a score, the benchmark it came from, what the leading model scored on the same test, and a link to the published result. Rows appear as results are published and read.

How we decide what counts as evidence

05 / capability

Capability shape

Where this model is strong, and against how many peers. Ranks are against models with evidence in that category, not against everything we track: ranking against models nobody tested would rank who published, not who is better.

No category scores to shape yet.

A category score is the weighted mean of the benchmarks published for it. With no published rows there is nothing to average, and an empty chart drawn at zero would say something false.

How we decide what counts as evidence

06 / cost

What it costs

List API rates as last read from the provider, with the source on every row, plus every change we have recorded since we started tracking it.

gpt-oss-20b API pricingSurge45°
ChargePriceUnitRead onSource
Input$0.02per 1M tokens2026-09-24Check
Output$0.09per 1M tokens2026-09-24Check

Where it is offered as a hosted endpoint, gpt-oss-20b sits at the very bottom of the market, cheap enough that per-call cost stops being a design constraint for most workloads. Self-hosting shifts the cost to hardware and operations instead, and the small active parameter count keeps that bill modest compared with running a dense model of similar total size.

What a month costsSurge45°
WorkloadInput / monthOutput / monthCost
A small product team20M tokens5M tokens$0.81
A busy support assistant200M tokens40M tokens$7.20
A document pipeline1000M tokens100M tokens$27.00

List API rates, no caching and no batch discount, which both providers offer and which change the answer a great deal. Treat these as the ceiling, not the bill.

07 / specs

Specifications

As listed by the provider's own catalogue and re-read every few hours. Anything absent is absent there too.

SpecificationSurge45°
Context window131,072 tokens
Maximum output32,768 tokens
Modalitiestext
Released05/08/2025
StatusCurrent
Catalogue identifieropenai/gpt-oss-20b

08 / lineage

Lineage

What this model replaced, what replaced it, and what else its provider has in the field.

Also from OpenAI

09 / line

The line

Every model in this family in release order, so a page from eight months ago says in one glance that two newer ones exist.

Came before

← gpt-oss-120b

Came after

Nothing newer in this line yet.

  1. 01gpt-oss-120b05/08/2025
  2. 02gpt-oss-20b05/08/2025

Ordered by release date and worked out from the naming, so a new member slots in as soon as its page exists. A retired model keeps its page and its place in the line.

10 / notes

Our notes

What this model changes for a brand trying to be cited in AI answers, and every change we have logged since it launched.

What it changes for you

Open weights mean more of the AI layer inside SaaS products is built rather than bought, so your brand is increasingly being summarised by models nobody chose for their knowledge of you. Small models like this lean hard on whatever text is put in front of them, which makes clean, retrievable, unambiguous product content matter more than authority signals. Write so that a model with a short attention span and no prior knowledge of your category can still describe you correctly from a single retrieved page.

Where buyers meet this model

Buyers rarely meet gpt-oss-20b as a branded product. They meet it inside other people's software: the summariser in a SaaS tool, an on-device or self-hosted assistant, a support triage layer, or a cheap tier on an inference platform. If a vendor tells you their AI features run on their own infrastructure, this is one of the models likely doing it.

Change log

Nothing published here yet. Changes appear within hours of a provider announcing them.

11 / questions

Questions

The things people ask about this model, answered from what is on this page rather than from anywhere else.

Can I use gpt-oss-20b commercially?
Yes. It is released under the Apache 2.0 licence, which permits commercial use, modification and redistribution without a separate agreement with OpenAI.
Does gpt-oss-20b handle images or audio?
No. It is a text-only model, so image, audio and video inputs need a different model in your stack.
How does it differ from OpenAI's hosted models?
You get the weights, so you can run it on your own infrastructure, fine-tune it and keep data inside your boundary. In exchange you give up the reasoning depth and the multimodal input of the larger hosted models.
What is it good for in a SaaS product?
High-volume routine text work: classification, extraction, summarising, tagging, drafting and simple tool-calling loops where the cost of a frontier model would not be justified.
Surge45°

Is gpt-oss-20b recommending you?

Models change what gets cited. We measure whether AI answers name your brand or your competitors across every assistant, and show you what to change.

Free AI visibility audit Book a call