benchmarks
GPT-OSS-120B is OpenAI’s open-weight mixture-of-experts text reasoning model, released under the Apache 2.0 license. It supports configurable reasoning effort, tool use and self-hosted deployment, with operating costs and safeguards determined by the hosting environment.
Benchmark results
Every figure links to where it was published. Compare like with like — harnesses and prompt setups differ between vendors.
| Benchmark | Score | Source |
|---|---|---|
| GPQA DiamondPhD-level science questions | 75.8%11 Dec 2025 | View |
| OTIS Mock AIME 2024-25Competition mathematics | 88.9%11 Dec 2025 | View |
| Terminal-BenchLong-horizon command-line agent tasks | 18.7%1 Nov 2025 | View |
| LMArena TextHuman preference Elo from blind chat votes (not a percentage) | 1366 Elo22 Oct 2026 | View |
| Aider PolyglotCode editing across six programming languages | 41.8% | View |
Weighted mean of 4 verified benchmark results: GPQA Diamond 75.8%, OTIS Mock AIME 2024-25 88.9%, Terminal-Bench 18.7%, Aider Polyglot 41.8%.
Capabilities
- Supports low, medium and high reasoning effort to balance response speed against reasoning depth.
- Supports function calling and tool-connected workflows, including browsing and code execution when tools are provided.
- Provides open weights under Apache 2.0 for self-hosting, modification and commercial deployment.
- Supports fine-tuning for specialized tasks and deployment requirements.
- Uses native MXFP4 quantization to enable documented deployment on a single 80 GB GPU.
What improved
- Compared with OpenAI’s earlier GPT-2 open-weight release, adds post-training for instruction following and reasoning.
- Compared with GPT-2, introduces configurable reasoning effort.
- Compared with GPT-2, supports structured tool-oriented conversations through the Harmony format.
- Compared with GPT-2’s 1,024-token context, expands the documented context window to 131,072 tokens.
Known flaws & limitations
- Text-only model; it does not natively process images, audio or video.
- Can hallucinate factual answers; OpenAI reports higher hallucination rates than some proprietary reasoning models.
- Correct integration requires support for Harmony’s conversation structure, roles, channels and tool conventions.
- The documented single-GPU deployment requires 80 GB of GPU memory, exceeding most consumer graphics cards.
- Tool-assisted benchmark results depend on the associated tools, execution environment and evaluation settings.
Risks, safety & security issues
- Modified or fine-tuned weights may change safety behavior; downstream deployments require independent safety evaluation.
- Tool-connected deployments can turn incorrect or adversarially influenced outputs into consequential actions unless permissions and execution are constrained.
- Plausible factual errors can mislead users in medical, legal, financial and other high-stakes applications.
Good fit for
- Self-hosted text assistants where organizations need control over model weights and infrastructure.
- Coding and software-maintenance assistants with sandboxed tools and human review.
- Mathematical and scientific reasoning workflows where outputs can be checked.
- Fine-tuned domain assistants with task-specific evaluation and safeguards.
Poor fit for
- Native image, audio or video understanding.
- Unsupervised high-stakes advice or decisions requiring consistently reliable factual accuracy.
- Unrestricted autonomous execution with access to sensitive systems or data.
- Local deployment on typical consumer GPUs without substantial deployment changes.
References
Profile last reviewed 10/27/2026 and re-checked automatically every few hours against the vendor's published documentation, papers and public leaderboards. Scores come from the cited benchmark results above — confirm limits and pricing against the provider's own documentation before building on them.