Can you own it?
This page is a projection of the one entry record, the Use & modify and Transparency factors that Assess covers. The full verdict is set by all four factors together, floor-weighted so the weakest caps the whole.
- AssessUse & modify + Transparency
- ImplementData control + Reliability
- UseReliability
- SupportTransparency
Intended & out-of-scope use
gpt-oss is OpenAI's first open-weight release since GPT-2 (August 2025). It comprises two mixture-of-experts reasoning models, gpt-oss-120b (~117B total / 5.1B active) and gpt-oss-20b (~21B total / 3.6B active), built for coding, math, agentic tool use and general reasoning with configurable low/medium/high effort. The intended use is general-purpose assistant and agentic workloads that need a strong reasoning model, run and customised on the deployer's own hardware under a permissive licence.
Only safety-tuned reasoning checkpoints ship (there is no raw base release), and the models are text-only. High-stakes, autonomous or regulated decision-making is out of scope without the external control stack described in the Implement domain. Because the weights are open, downstream parties can fine-tune the safety tuning away.
Known limitations, bias & failure modes
Strong reasoning does not remove confabulation, so factual output should be treated as unverified.
Raw chain-of-thought is unfiltered. OpenAI deliberately did not align the visible reasoning trace, so it can contain unsafe or incorrect content and must not be shown to end users or used as a safety signal.
The model card presents training as primarily English, so other-language behaviour is incidental and not a supported target.
Safety tuning is removable. Because the weights are open, the safety behaviour can be fine-tuned out, which is the risk OpenAI probed with its malicious-fine-tuning study.
The models are trained solely for the harmony response format, so plain prompt strings underperform or misbehave.
Openness tier & components
gpt-oss is open_weights, not fully_open. Open: ungated, downloadable Apache-2.0 weights and
strong public documentation (a detailed model card and arXiv report). Closed: the training
data and training code. Evaluation is only partially open.
The release is more open than a gated community licence, but it does not publish the data and code needed to reproduce the model from scratch. The open-weights tier caps Dimension 1 at 3, and gpt-oss sits at the top of that band.
License terms & permitted use
The licence is Apache-2.0, OSI-approved, applied to both models with no click-through gate. There is no acceptable-use policy, no monthly-active-user threshold, no attribution or naming requirement and no field-of-use restriction. Commercial use is unconditional, subject only to the standard Apache attribution and notice terms.
This is among the least restrictive licences in the registry and feeds both the openness (3) and legal (4) scores. The legal score is held at 4 by the EU AI Act training-content-summary gap (see below), not by anything in the licence.
Supply-chain & provenance
Weights are distributed from the verified openai org on Hugging Face as safetensors (no
pickle requirement) with per-file checksums, shipped natively in MXFP4. There is a clear
canonical source and no malicious-checkpoint incident on record for the canonical org. The
checkpoint trust checklist scores 6/8. The two missing controls are cryptographic weight
signing (Sigstore/model-signing) and SLSA build attestation, which is why provenance scores 4,
not 5.
Community MXFP4/GGUF redistributions (Ollama, LM Studio, llama.cpp) are separate artifacts.
Pin the exact openai revision and verify checksums.
EU AI Act posture
gpt-oss is a GPAI model under the 10²⁵-FLOPs systemic-risk threshold (gpt-oss-120b was trained on roughly 2.1M H100-hours, an estimated <1e25 FLOPs), so no Article 55 regime applies. Released under an OSI-approved licence, ungated and not monetised, it plausibly qualifies for the Article 53 open-source exemption from the Annex XI/XII technical-documentation duties.
The surviving obligation is a public training-content summary. gpt-oss does not publish its training data, so that summary depends on OpenAI authoring it and cannot be reconstructed from a released corpus. This is why the legal score is capped at 4.
OpenAI has committed to the EU GPAI Code of Practice. A fine-tuner placing a derivative on the EU market under its own name may become a provider for that derivative and must maintain its own documentation.
Benchmarks & evaluation
Aggregated from OpenAI's model card and independent leaderboards (ev-perf,
ev-model-card): gpt-oss-120b is reported near OpenAI o4-mini and gpt-oss-20b near o3-mini on
core reasoning benchmarks, and both rank strongly among open-weight models on public
leaderboards. The scores are publisher and third-party framing, so the performance dimension
is capped at 4.
Independent safety evaluation
The safety evaluation is a strength relative to most open-weight families, but it is partial. OpenAI published a substantial safety evaluation under its Preparedness Framework, including a worst-case malicious-fine-tuning study that fine-tuned gpt-oss on biology and cyber data to probe uplift. It concluded that the models did not reach "High" capability thresholds.
The methodology was independently reviewed by external experts (METR, SecureBio and Daniel
Kang; ev-metr), which is why misuse_exposure is assessed partial and the safety dimension
can cite a third-party source. It stops short of a full independent red-team across all harm
domains, so coverage is marked partial.
How this scores
The ownership factors this domain covers, drawn from the one entry record.
Use and modify freelyCan you run, modify and adapt it with no gate and no field-of-use trap?
StrongThe Apache-2.0 grant is OSI-approved: the weights can be owned, run, modified and redistributed freely, with no access gate and no field-of-use limit. The only usage constraint is a two-sentence policy asking users to obey applicable law, an unusually light AUP for an open-weight release.
TransparencyDo you know what it is: weights, training, behaviour, and legible terms?
ModerateWeights are open and inspectable, and OpenAI published a safety report whose worst-case malicious-fine-tuning methodology was independently reviewed by METR. Training data and pre-training code are closed, so the weights can be adapted but not reproduced, and how the model was made cannot be seen.
Sources
The same evidence records as the entry sheet. Read means the text was verified; unverified means it is known to exist but not yet read.