gpt-oss-120b and gpt-oss-20b are open-weight reasoning models from OpenAI, released on 5 August 2025 under the Apache 2.0 licence. This profile collects the published facts about both models in one place: parameter counts, architecture, context window, output limits, knowledge cutoff, hardware needs, tools, benchmarks from the model card, hosted availability on Amazon Bedrock and lifecycle status. It is written for developers, platform teams, buyers and governance teams deciding whether to self-host or call a hosted endpoint. Every fact comes from OpenAI's own documentation, model card and repository, or from AWS documentation for Bedrock details, listed under Standards and references. Where a provider publishes nothing on a point, the page says so.
Last checked: 2026-11-15.
This is an independent reference page from NHI Mgmt Group based on public provider documentation; it is not sponsored or endorsed by any provider. It uses OpenAI's public documentation as of November 2026. Model details change, so confirm against the provider's current documentation.
Key takeaways
- Released 5 August 2025, Apache 2.0. OpenAI's launch post and model card give the release date; weights, inference code, tool environments and tokenizer were published under Apache 2.0.
- 117B and 21B total parameters, mixture of experts. gpt-oss-120b activates 5.1B parameters per token and gpt-oss-20b activates 3.6B, according to OpenAI's launch post and model pages.
- 131,072-token context on OpenAI's pages, 128K on Bedrock. OpenAI's model pages list 131,072 tokens for context and output; the AWS Bedrock model cards list a 128K context and a 16K maximum output.
- Runs on 80 GB and 16 GB of memory. OpenAI states gpt-oss-120b runs within 80 GB (one H100 class GPU) and gpt-oss-20b within 16 GB, using native MXFP4 quantisation.
- No per-token price from OpenAI. OpenAI's model pages show no price and zero API rate limits for the Build, Launch and Grow tiers; hosted access is listed through Amazon Bedrock, which points to its own pricing page.
What is gpt-oss-120b and gpt-oss-20b?
gpt-oss is OpenAI's open-weight model family. The two models share an architecture and training recipe and differ in size. OpenAI's model page describes gpt-oss-120b as its "most powerful open-weight model" that "fits into an H100 GPU", and describes gpt-oss-20b as a "medium-sized open-weight model for low latency" aimed at local or specialised use. Both accept text and return text, and both are reasoning models that produce a chain of thought before the answer.
The model card (arXiv 2508.10925, submitted 8 August 2025) describes both as mixture-of-experts transformers trained with large-scale distillation and reinforcement learning, tuned for agentic work: browsing, running Python and calling developer-defined functions. They are post-trained on OpenAI's harmony response format, and the repository states they "should only be used with this format". For OpenAI's full line-up, see OpenAI Models. For how they compare with other open-weight families, see Open-Weight Frontier Models Compared.
Key facts at a glance
As of 2026-11-15. Sources: OpenAI model pages, launch post, model card and AWS Bedrock model cards.
| Field | Value | Source |
|---|---|---|
| Provider | OpenAI | OpenAI model pages; launch post |
| Model IDs | gpt-oss-120b; gpt-oss-20b (no dated snapshots listed) | OpenAI model pages |
| Release date | 2025-08-05 | Launch post; model card page |
| Status | Current open-weight models; no deprecation notice | OpenAI model pages; deprecations page |
| Total parameters | 120b: 117B; 20b: 21B | Launch post; model pages |
| Active parameters per token | 120b: 5.1B; 20b: 3.6B | Launch post; model pages |
| Layers and experts | 120b: 36 layers, 128 experts, 4 active; 20b: 24 layers, 32 experts, 4 active | Launch post |
| Context window | 131,072 tokens (OpenAI); 128K (Bedrock) | OpenAI model pages; Bedrock model cards |
| Max output | 131,072 tokens (OpenAI); 16K (Bedrock) | OpenAI model pages; Bedrock model cards |
| Knowledge cutoff | 1 June 2024 | OpenAI model pages; model card |
| Input modalities | Text | OpenAI model pages |
| Output modalities | Text | OpenAI model pages |
| Reasoning effort | low, medium, high | OpenAI model pages |
| Open weights | Yes, on Hugging Face | OpenAI model pages; Hugging Face model card |
| Licence | Apache 2.0, plus a USAGE_POLICY file in the repository | Model pages; GitHub repository |
| Fine-tuning | Supported (parameter fine-tuning); 120b on a single H100 node | OpenAI model pages; Hugging Face model card |
| Preparedness rating | Default 120b below High in Biological and Chemical, Cyber and AI Self-Improvement | Model card page |
Release and version history
As of 2026-11-15. Dates from OpenAI's launch posts, arXiv and AWS documentation.
| Date | Event | Source |
|---|---|---|
| 2025-08-05 | gpt-oss-120b and gpt-oss-20b released with weights on Hugging Face; model card published; Red Teaming Challenge announced with a $500,000 prize fund | Launch post; model card page |
| 2025-08-05 | Both models launched on Amazon Bedrock | Bedrock model cards |
| 2025-08-08 | Model card submitted to arXiv (2508.10925v1) | arXiv |
| 2025-10-29 | gpt-oss-safeguard-120b and gpt-oss-safeguard-20b released as fine-tuned versions for policy-based safety classification (research preview, Apache 2.0) | gpt-oss-safeguard launch post |
OpenAI lists no dated snapshots on either model page, so the weights published at launch are the reference version. The safeguard models are separate classifiers and are not covered in detail here.
How much does gpt-oss-120b and gpt-oss-20b cost?
As of 2026-11-15. The weights are free to download under Apache 2.0. Running them costs whatever the chosen hardware or hosting service charges.
| Item | Price | Notes | Source |
|---|---|---|---|
| Weights download | No charge | Apache 2.0, Hugging Face | OpenAI model pages; Hugging Face model card |
| OpenAI API, per token | Not stated by OpenAI on the pages checked (2026-11-15) | Model pages show zero rate limits for Build, Launch and Grow; Free tier not supported | OpenAI model pages |
| Amazon Bedrock, per token | Not stated on the Bedrock model cards checked (2026-11-15) | Model cards refer to the Amazon Bedrock pricing page; Standard, Priority and Flex tiers supported, Reserved not supported | Bedrock model cards |
| Self-hosting hardware | Not a price: 80 GB GPU memory (120b), 16 GB (20b) | Single H100 or AMD MI300X named for 120b | Launch post; Hugging Face model card |
Other hosts named in OpenAI's launch post set their own prices. Compare hosted rates in the LLM Pricing Tracker and confirm on each host's own pricing page.
Context window and output limits
The model card states that dense attention layers were extended to 131,072 tokens using YaRN. The figures differ by channel, so check the limit of the endpoint you actually call.
| Limit | Value | Notes | Source |
|---|---|---|---|
| Context window (OpenAI) | 131,072 tokens | Both sizes | OpenAI model pages; model card |
| Max output (OpenAI) | 131,072 tokens | Both sizes; reasoning tokens supported | OpenAI model pages |
| Context window (Bedrock) | 128K tokens | Both sizes | Bedrock model cards |
| Max output (Bedrock) | 16K tokens | Both sizes | Bedrock model cards |
| Attention pattern | Alternating banded window (128-token bandwidth) and dense layers | 64 query heads, grouped query attention with 8 key-value heads | Model card |
| Tokenizer | o200k_harmony, 201,088 tokens | Open-sourced with the harmony renderer | Model card; launch post |
| OpenAI API rate limits | 0 RPM, 0 TPM (Build, Launch, Grow) | Free tier not supported | OpenAI model pages |
Modalities and tools
Both models accept text and return text. OpenAI's model pages mark image, audio and video as not supported.
- Features: streaming, function calling and Structured Outputs are listed as supported; OpenAI's launch post describes function calling as few-shot.
- Built-in tools: web browsing and Python code execution, trained into the model through the harmony format. The repository ships a reference browser tool (search, open, find) and a stateless Python tool.
- Reasoning effort: low, medium and high, set with one line in the system message (for example "Reasoning: high"), according to the launch post and Hugging Face card.
- Chain of thought: full chain of thought is available. OpenAI states it applied no direct supervision to it so that developers can monitor it, and the Hugging Face card says it is not meant for end users.
- Recommended sampling: temperature 1.0 and top_p 1.0 (repository).
- Precision: MoE weights in MXFP4, other tensors in BF16; BF16 recommended for activations.
- Inference runtimes named: Transformers, vLLM, SGLang, Ollama, LM Studio, llama.cpp quantisations, and OpenAI's PyTorch, Triton and Metal reference implementations (Metal marked not production-ready).
The repository warns that the reference Python tool "runs in a permissive Docker container which could be problematic in cases like prompt injections", and that the browser tool is for educational use only. Build your own sandbox before giving either model tools. See the Open-Weight Model Security Guide and MITRE ATT&CK T1698 Indirect Prompt Injection.
Where is gpt-oss-120b and gpt-oss-20b available?
As of 2026-11-15.
| Channel | Details | Source |
|---|---|---|
| Hugging Face (weights) | openai/gpt-oss-120b and openai/gpt-oss-20b; also HuggingChat and listed inference providers | Hugging Face model card; OpenAI model pages |
| OpenAI API | Model pages list endpoints, but rate limits are 0 for Build, Launch and Grow and Free tier is not supported | OpenAI model pages |
| Amazon Bedrock | bedrock-runtime IDs openai.gpt-oss-120b-1:0 and openai.gpt-oss-20b-1:0 (Chat Completions, Converse, Invoke); bedrock-mantle IDs openai.gpt-oss-120b and openai.gpt-oss-20b (Responses, Chat Completions); 17 in-Region locations including us-east-1, eu-west-2 and us-gov-west-1; GovCloud geo inference only; global cross-Region not supported | Bedrock model cards |
| Other partners at launch | Azure, AWS, Fireworks, Together AI, Baseten, Databricks, Vercel, Cloudflare, OpenRouter, Groq, Cerebras and others; gpt-oss-20b on Windows via ONNX Runtime | Launch post |
| ChatGPT plans | Not stated by OpenAI on the pages checked (2026-11-15) | OpenAI model pages |
The partner list is OpenAI's launch-day list; current availability on each platform should be confirmed with that platform. On Bedrock, the runtime endpoint supports Guardrails and streaming but not agents, knowledge bases or abuse detection, according to the 120b model card. See the Amazon Bedrock Security Guide.
Access restrictions
There is no gated access: the weights download from Hugging Face without an approval step on the pages checked, and Apache 2.0 permits commercial use, modification and redistribution. The repository includes a USAGE_POLICY file alongside the licence. Once weights are released, OpenAI's model card page notes that it cannot add mitigations or revoke access, and that open-weight models can be fine-tuned to bypass refusals. Hosted channels apply their own account controls: Bedrock requires an AWS account with model access in a supported Region. The OpenAI API rate limits of zero mean the hosted OpenAI route is not a practical access path on the pages checked.
Benchmarks (provider-reported, dated)
As of 2026-11-15. Provider-reported figures from Table 3 of the gpt-oss model card (arXiv 2508.10925v1, 8 August 2025), at high reasoning effort. Independent leaderboard results are not included because none were checked against their own published pages for this profile.
| Evaluation | gpt-oss-120b | gpt-oss-20b | Source and date |
|---|---|---|---|
| AIME 2024 (no tools) | 95.8 | 92.1 | Model card, 2025-08-08 |
| AIME 2025 (with tools) | 97.9 | 98.7 | Model card |
| GPQA Diamond (no tools) | 80.1 | 71.5 | Model card |
| MMLU | 90.0 | 85.3 | Model card |
| Humanity's Last Exam (no tools) | 14.9 | 10.9 | Model card |
| Codeforces Elo (no tools) | 2463 | 2230 | Model card |
| SWE-bench Verified | 62.4 | 60.7 | Model card |
| Tau-Bench Retail | 67.8 | 54.8 | Model card |
| HealthBench | 57.6 | 42.5 | Model card |
| HealthBench Hard | 30.0 | 10.8 | Model card |
| SimpleQA hallucination rate (no browsing) | 0.782 | 0.914 | Model card |
| PersonQA hallucination rate (no browsing) | 0.491 | 0.532 | Model card |
For hallucination rates lower is better; for the other scores higher is better. The model card notes both models underperform OpenAI o4-mini on SimpleQA and PersonQA, which it attributes to smaller models holding less world knowledge. Plan retrieval or browsing for factual tasks.
Model card and system card
OpenAI published a combined "gpt-oss-120b & gpt-oss-20b Model Card" on 5 August 2025, also on arXiv. There is no separate system card. What the card and its summary page state:
- Preparedness: default gpt-oss-120b did not reach the indicative High thresholds in Biological and Chemical, Cyber or AI Self-Improvement.
- Adversarial fine-tuning: OpenAI fine-tuned the model on biology and cybersecurity data to simulate a determined attacker; its Safety Advisory Group concluded it still did not reach High in Biological and Chemical or Cyber.
- Frontier comparison: for most biology evaluations, existing open models nearly match the adversarially fine-tuned gpt-oss-120b, so OpenAI concludes the release does not significantly advance the open-model frontier there.
- Training safeguards: certain harmful CBRN data was filtered during pre-training; post-training used deliberative alignment and the instruction hierarchy to refuse unsafe prompts and resist prompt injection (launch post).
- Deployer responsibility: developers may need extra safeguards to match the protections of OpenAI's hosted models.
For how to read these documents, see Model Cards and System Cards Explained and the Frontier AI Model Risk Frameworks Guide.
Data handling
Data handling depends on where the model runs, because OpenAI does not operate the inference when the weights are self-hosted.
- Self-hosted: prompts, outputs and chain of thought stay in the deployer's own infrastructure, and retention, logging and access are the deployer's responsibility. OpenAI publishes no data policy specific to gpt-oss on the pages checked.
- Amazon Bedrock: AWS states that model providers "don't have access to Amazon Bedrock logs or to customer prompts and completions". Retention and training-use terms are on AWS pages not checked for this profile.
- Chain of thought: because it is unfiltered, treat stored reasoning traces as sensitive data in the same class as the prompt, and do not show them to end users.
- Tool traffic: the browser tool sends queries to whichever search backend the deployer configures, so that backend's terms apply.
Self-hosting moves secrets to your side: model registry tokens, inference endpoint keys and the credentials tools use are non-human identities. See the Hugging Face Security Guide.
Deprecation status
As of 15 November 2026, OpenAI's deprecations page lists no entry for gpt-oss-120b or gpt-oss-20b (the fetched page mentions gpt-oss only in navigation links; a small tail of the page was not read). Released weights cannot be withdrawn from copies already downloaded, so self-hosted deployments are governed by the deployer's own lifecycle. On Bedrock, AWS lists both models as Active, with end of life no sooner than 5 August 2026 and a legacy period of at least six months; that date has passed, and the model EOL date is shown as N/A.
| Model or item | Details | Source |
|---|---|---|
| gpt-oss-120b (OpenAI) | No deprecation entry found (2026-11-15) | OpenAI deprecations page |
| gpt-oss-20b (OpenAI) | No deprecation entry found (2026-11-15) | OpenAI deprecations page |
| Bedrock gpt-oss-120b and gpt-oss-20b | Active; EOL no sooner than 5 August 2026; legacy period at least 6 months; model EOL date N/A | Bedrock model cards |
How gpt-oss-120b and gpt-oss-20b differ from other models in their family
As of 2026-11-15. Models listed by size. Sources: launch posts and OpenAI model pages.
| Model | Released | Total / active parameters | Memory stated | Purpose (OpenAI wording) |
|---|---|---|---|---|
| gpt-oss-120b | 2025-08-05 | 117B / 5.1B | 80 GB | Most powerful open-weight model, fits into an H100 GPU |
| gpt-oss-20b | 2025-08-05 | 21B / 3.6B | 16 GB | Medium-sized open-weight model for low latency |
| gpt-oss-safeguard-120b and -20b | 2025-10-29 | Not stated in the launch post checked | Not stated in the launch post checked | Fine-tuned gpt-oss for safety classification against a developer-supplied policy |
Both main models share the knowledge cutoff, context length, tokenizer, tools and licence. The documented differences are size, memory, and benchmark results: the larger model leads on most rows of the model card table, while gpt-oss-20b scores slightly higher on AIME 2025 with tools. Unlike OpenAI's GPT-6 models, neither is offered with a published per-token price or rate limit on OpenAI's API. For a cross-provider view, see gpt-oss vs Gemma 4 vs Mistral Small 4 and the Open-Weight Model Licence Tracker.
Practitioner checklist
- Record which weights you run (Hugging Face repository and commit) and verify file hashes before deployment.
- Serve the models with the harmony format through a supported runtime; raw prompts without it give unreliable output.
- Confirm the context and output limits of your chosen endpoint: 131,072 tokens on OpenAI's pages, 128K and 16K on Bedrock.
- Run the Python and browser tools in your own locked-down sandbox with no ambient credentials and restricted egress.
- Keep chain of thought out of user interfaces and treat logged reasoning traces as sensitive data.
- Add your own input and output safety layers, since OpenAI cannot update or revoke released weights.
- Use retrieval or browsing for factual answers, given the model card's SimpleQA and PersonQA hallucination rates.
- Store registry tokens and inference endpoint keys in a secrets manager and scope them per environment.
- Check the Bedrock lifecycle fields and OpenAI's deprecations page when planning multi-year use.
Standards and references
- OpenAI: gpt-oss-120b model page (accessed 2026-11-15)
- OpenAI: gpt-oss-20b model page (accessed 2026-11-15)
- OpenAI: Introducing gpt-oss (accessed 2026-11-15)
- OpenAI: gpt-oss-120b and gpt-oss-20b Model Card (arXiv 2508.10925) (accessed 2026-11-15)
- OpenAI: gpt-oss model card summary (accessed 2026-11-15)
- OpenAI: gpt-oss GitHub repository (accessed 2026-11-15)
- Hugging Face: openai/gpt-oss-120b model card (accessed 2026-11-15)
- AWS: Amazon Bedrock gpt-oss-120b model card (accessed 2026-11-15)
- AWS: Amazon Bedrock gpt-oss-20b model card (accessed 2026-11-15)
- OpenAI: Deprecations (accessed 2026-11-15)
Related NHI Mgmt Group resources: OpenAI Models, Open-Weight Frontier Models Compared, gpt-oss vs Gemma 4 vs Mistral Small 4, Open-Weight Model Security Guide, Amazon Bedrock Security Guide.