gpt-oss-20b
gpt-oss-20b is the smaller of OpenAI's two open-weight models (Apache 2.0, August 2025): a mixture-of-experts network with 21B parameters of which 3.6B are active per token. OpenAI positions it for low-latency and specialised use — function calling, structured outputs, agentic steps — with reasoning effort set to low, medium or high. A sensible default when the task is routine and the cost per call matters. It is fulfilled through global providers under our European contract, and requests go only to backends that do not retain prompts or completions.
model id: gpt-oss-20b
from openai import OpenAI client = OpenAI( api_key="hb-...", base_url="https://api.heabsy.com/v1",) r = client.chat.completions.create( model="gpt-oss-20b", messages=[{"role": "user", "content": "Hello"}],)print(r.choices[0].message.content)
curl https://api.heabsy.com/v1/chat/completions \ -H "Authorization: Bearer $HEABSY_API_KEY" \ -H "Content-Type: application/json" \ -d '{ "model": "gpt-oss-20b", "messages": [{"role": "user", "content": "Hello"}] }'
import OpenAI from "openai"; const client = new OpenAI({ apiKey: process.env.HEABSY_API_KEY, baseURL: "https://api.heabsy.com/v1",}); const r = await client.chat.completions.create({ model: "gpt-oss-20b", messages: [{ role: "user", content: "Hello" }],});console.log(r.choices[0].message.content);
Price list, updated 12 September 2026. No minimum spend, no subscription.
About the model
- Developer
- OpenAI
- Released
- August 2025
- Licence
- Apache-2.0
- Architecture
- mixture of experts · 21B parameters, 3.6B active per token
- Input
- text
- Reasoning
- selectable effort
- Knowledge cutoff
- June 2024
- Languages named by the developer
- none named on the model card
Source: the developer's model card, as of 17 September 2026.
gpt-oss-120b or gpt-oss-20b?
Both are OpenAI's open-weight models from August 2025 under Apache 2.0, both text-only, both with three reasoning levels (low, medium, high). They are not the models behind ChatGPT.
gpt-oss-120b has 117B parameters with 5.1B active per token and is aimed at production reasoning; gpt-oss-20b has 21B with 3.6B active and fits into 16 GB of memory, which makes it the choice for routine calls where latency and price decide.
OpenAI does not list supported languages but measured 14 of them on the multilingual MMMLU test (at high reasoning effort, German: 83.0 for the 120b, 78.7 for the 20b). OpenAI gives a knowledge cutoff of June 2024. If you run the weights yourself, the card requires OpenAI's harmony response format; through our API you use the ordinary chat completions call shown above.
Licence: what it means for your company
Apache-2.0. A permissive open-source licence: you may use, modify and sell products built on the model, including commercially; when you redistribute the weights, the licence and copyright notices go with them.
A summary of the licence, not legal advice.
Benchmarks reported by the developer
- AIME 2025
- 91.7 %
- GPQA Diamond
- 71.5 %
- MMLU
- 85.3 %
- SWE-bench Verified
- 60.7 %
Figures from the developer's model card, not our measurement. Source: arxiv.org/abs/2508.10925
Technical details
- Context window
- 128K tokens
- Max output
- 65,536 tokens
- Throughput limit
- 200K tok/min
- Tokenizer
- Other
- Streaming
- yes
- Open weights
- openai/gpt-oss-20b
- max_tokens
- 1 … 65536
- temperature
- 0 … 2
- top_p
- 0 … 1
- frequency_penalty
- -2 … 2
- presence_penalty
- -2 … 2
- stop
- array
- seed
- integer
This model is fulfilled through global providers under our European contract: one DPA with an EU company, one invoice for every model in the catalog, no US counterparty on your paperwork. Compute may run outside the EEA — if you need strict EEA-only processing, use our own-hardware model.
Questions about this model
›Where is gpt-oss-20b computed?
Outside the EEA, through a global provider under our European contract. You get one DPA with an EU company and one invoice, but the compute itself does not run in the EEA — we state that instead of hiding it.
›Does gpt-oss-20b store prompts and completions?
No. This model runs with zero data retention: request content is not written to durable storage and nothing is used for training.
›How large is the gpt-oss-20b context window?
128K tokens, with up to 65,536 tokens of output.
›What does gpt-oss-20b cost?
$0.05 per million input tokens and $0.20 per million output tokens. No minimum spend and no subscription; the same price list covers the API and the invoice.
›Is there a throughput limit on gpt-oss-20b?
Yes, 200K tokens per minute. Higher limits are a question of volume, not of principle.
›Which languages does gpt-oss-20b support?
The developer's model card names no languages. Test the model on your own material before relying on it in a particular language. OpenAI did measure 14 languages on the multilingual MMMLU test, German among them.
›Under which licence is gpt-oss-20b released?
Apache-2.0. A permissive open-source licence: you may use, modify and sell products built on the model, including commercially; when you redistribute the weights, the licence and copyright notices go with them.
›Can reasoning be switched off in gpt-oss-20b?
The developer lets you choose how much the model reasons; the card does not document turning reasoning off completely.
Compare with similar models
| Model | Developer | Parameters | Context | Licence | Input / 1M |
|---|---|---|---|---|---|
| gpt-oss-20b | OpenAI | 21B / 3.6B | 128K | Apache-2.0 | $0.05 |
| gpt-oss-120b | OpenAI | 117B / 5.1B | 128K | Apache-2.0 | $0.11 |
| Mistral Small 24B (2501) | Mistral AI | 24B | 32K | Apache-2.0 | $0.05 |
| Cydonia 24B v4.1 | TheDrummer | 24B | 128K | — | $0.42 |
30 more models behind the same key and the same invoice.
Browse the catalog