DeepInfra has announced the immediate availability of two new open-weight models from OpenAI, gpt-oss-20B and gpt-oss-120B, on its platform. This release marks a significant development in the accessibility of advanced AI models, offering developers powerful new tools for various applications. The models are designed for agentic tasks, high reasoning, and versatile developer use cases.
The gpt-oss-20B model is priced at $0.04 per million input tokens and $0.16 per million output tokens on DeepInfra. Its larger counterpart, gpt-oss-120B, is available for $0.09 per million input tokens and $0.45 per million output tokens. As DeepInfra stated in a tweet, these models are
"Agentic, fast, open-source. As always - best price."
OpenAI's gpt-oss-120B model achieves near-parity with the company's proprietary o4-mini on core reasoning benchmarks. The gpt-oss-20B model delivers comparable results to OpenAI's o3-mini, making it suitable for edge devices and local inference. Both models utilize a Mixture-of-Experts (MoE) architecture, with gpt-oss-120B having 117 billion parameters and gpt-oss-20B featuring 21 billion parameters.
Crucially, these gpt-oss models are released under the permissive Apache 2.0 license, allowing for broad commercial use, modification, and redistribution. This move represents OpenAI's first open-weight model release since GPT-2 in 2019, signaling a strategic shift towards greater openness in the AI ecosystem. The Apache 2.0 license provides developers and enterprises with unprecedented control over deployment and customization.
Beyond DeepInfra, the gpt-oss models are also available on platforms like Hugging Face, Azure AI Foundry, and Amazon Bedrock, broadening their accessibility. This widespread availability underscores a growing industry trend towards open-weight models, empowering a wider range of developers to build and innovate with advanced AI. The models are designed to run efficiently, with gpt-oss-120B capable of running on a single 80 GB GPU.