gpt-oss family

Open-weight general-purpose reasoning model family released for self-hosting and third-party deployment.

Score calculation

(4 + 8 + 8 + 4 + 5) ÷ 5 = 5.8

5 of six dimensions rated. Unrated dimensions and BDS status are excluded.

Setup details

This entry covers gpt-oss-120b and gpt-oss-20b core reasoning models. OpenAI separately describes gpt-oss-safeguard as research-preview safety reasoning models. Privacy depends on the chosen host; OpenAI company policies and contracts are not automatically attributed to these exact weights.

OpenAI’s current open-weight model page, updated October 19, 2026, lists gpt-oss-120b and gpt-oss-20b; the release announcement dates the family to August 5, 2025. The checked catalog provides no newer core gpt-oss family release.

Energy

↑

OpenAI reports training compute of 2.1 million H100-hours for gpt-oss-120b and nearly ten times fewer hours for gpt-oss-20b. GPU-hours are not facility electricity or emissions; the checked sources give no measured kWh, water, or standardized inference-energy figure.

Reasoning & sources

OpenAI reports H100 training-hours for both sizes, which gives an auditable compute quantity but not electricity or emissions. Hardware utilization and data-center energy boundaries are still needed; accelerator-hours alone cannot be converted to kWh.

No measured electricity, carbon intensity, water, or per-request inference energy is reported.

OpenAI deployment safety: gpt-oss ↗

Training rights

↑

OpenAI describes training on mostly English text covering STEM, coding, and general knowledge, but does not provide a source-level training-corpus manifest or independent rights audit. The Apache 2.0 license applies to the weights and does not establish training-data rights.

Reasoning & sources

OpenAI describes broad training categories but not a source-level corpus or rights audit; Apache 2.0 applies to the weights, not the training data. These materials do not establish permission for every source, so users should request provenance before making rights claims.

Broad training-data categories are described, while provenance and rights clearance remain opaque.

Source-level corpus, consent/licensing evidence, and independent rights review were not found.

Privacy

↑

OpenAI says the weights can be run on user-managed or hosting-provider infrastructure and that it does not receive self-hosted inputs unless they are explicitly shared. Privacy therefore depends on the operator and any managed host.

Reasoning & sources

Self-hosted weights can keep prompt data outside OpenAI’s service, according to its deployment guidance, but the operator then controls logging and access. Self-hosting can protect prompts from OpenAI’s service, but only with sound local controls; managed hosting needs a separate contract review.

Self-hosting can keep prompts outside OpenAI’s service, although the operator’s safeguards determine actual privacy.

The model does not itself enforce retention, access controls, telemetry practices, or host-side data handling.

Openness

↑

The gpt-oss-120b and 20b weights are downloadable under Apache 2.0, with model cards and safety documentation. OpenAI also describes gpt-oss-safeguard research-preview variants. Training data and the full training pipeline are not released.

Reasoning & sources

Downloadable Apache 2.0 weights and public model/safety cards support broad adaptation and inspection. The missing source corpus and full training pipeline prevent full reproducibility, so openness is substantial but bounded.

Permissively licensed weights and substantial documentation are available, with training-data and pipeline limits.

No source-level training corpus or fully reproducible training pipeline is public; open weights do not imply open surrounding services.

Worker treatment

↑

OpenAI’s supplier code sets broad labor standards but does not publish worker outcomes. TIME’s investigation reviewed OpenAI–Sama contracts, payslips, and interviews about 2021–22 Kenyan toxic-content labeling; workers described low take-home pay and distress, while Sama disputed workload/counseling claims and OpenAI said Sama managed pay and support. TIME later reported OpenAI said it was unaware of the conditions and no longer worked with Sama; this is a historical contractor case, not evidence of current API annotation conditions.

Reasoning & sources

A documented past supplier episode includes worker reports of low pay and psychological distress, countered by Sama and OpenAI responses; the current supplier code has no public independent outcome audit. This is relevant when deciding whose business to support, while pay and working conditions for the team behind gpt-oss family remain unestablished. Ask for current worker outcomes for that team and its suppliers.

The release and model card do not report annotation workers’ pay, contractors, or working conditions. Safety documentation and model weights do not answer labor questions; seek workforce sourcing details before drawing a labor conclusion.

A documented past supplier episode includes worker reports of low pay and psychological distress, countered by Sama and OpenAI responses; the current supplier code has no public independent outcome audit. This rating describes the company relationship or supplier policy, not a finding about this particular product’s use.

Working conditions for the people building or supporting gpt-oss family are not established by this company-wide record. The sources do not establish whether OpenAI uses other current annotation vendors, present worker outcomes, or the effectiveness of the supplier code across vendors.

Government & military

↑

OpenAI publicly describes a 2026 classified Department of War agreement with cloud-only deployment, retained safety controls, cleared staff, and contractual limits. It separately says ChatGPT will reach GenAI.mil for unclassified work used by 3 million civilian and military personnel. Both are direct defense deployments; limits are company-reported and not independently audited.

Reasoning & sources

OpenAI reports cloud-only deployment, retained safety controls, cleared staff, and contractual limits; these are meaningful safeguards, but no independent check of enforcement is public. This is relevant when deciding whose business to support. Deployment of gpt-oss family or this particular model release under those contracts remains unestablished; consider the company relationship separately from the product’s actual use.

OpenAI’s Department of War agreement is a company-level relationship, while the checked public description does not name gpt-oss. The agreement does not establish military use of these downloadable weights; users with defense restrictions should seek an explicit model-specific deployment statement.

OpenAI reports cloud-only deployment, retained safety controls, cleared staff, and contractual limits; these are meaningful safeguards, but no independent check of enforcement is public. This rating describes the company relationship or supplier policy, not a finding about this particular product’s use.

Deployment of gpt-oss family under these contracts is not established. Actual use, oversight effectiveness, and compliance with contract limits are not independently established.

BDS status

↑

We did not find OpenAI among the named targets in the two official BNC campaign lists. This is a limited list check, not BDS approval.

Reasoning & sources

OpenAI is not a named target in the checked lists, but BDS uses selective campaigns, so absence cannot establish a clean record. OpenAI’s April 2026 agreement names Microsoft as its primary cloud partner and a major shareholder, with revenue-share payments to Microsoft through 2030. That is a documented commercial link to a BDS target, not a separate official listing of OpenAI; local use of downloaded weights does not by itself establish a new payment to either company.

This review covers the named targets in the BNC consumer/pressure guide and No Tech campaign, not every BDS campaign, local group, investment or supplier. Lists and ownership can change. A company’s absence from these lists is not an endorsement or proof that it has no relevant relationships.

All sources