OpenAI announced on 5 August 2025 the release of gpt-oss, a new model described as optimised for open-source software development tasks [source]. The provider states the model is trained on public code repositories and designed to assist with code generation, debugging, and documentation.

According to the changelog, gpt-oss is available through the standard API with a dedicated model identifier. OpenAI reports the model demonstrates improved performance on benchmarks including HumanEval and MBPP compared to prior GPT-4 variants, though specific numerical comparisons were not provided in the announcement.

The provider notes gpt-oss incorporates "enhanced context handling for large codebases" and supports multiple programming languages. No details were given regarding training data cutoff dates, parameter count, or architectural modifications relative to existing models.

OpenAI states the model is subject to the same usage policies as other GPT models, including restrictions on generating malicious code. The announcement does not address whether the model itself is open-source or open-weight—the naming refers to its intended use case rather than licensing terms.

The release follows similar code-focused model announcements from other providers in recent months. OpenAI indicates gpt-oss will be priced identically to GPT-4 Turbo for input and output tokens.

No independent benchmarks or third-party evaluations were available at the time of announcement. The provider has not disclosed whether existing GPT-4 users will experience automatic routing to gpt-oss for code-related queries or whether explicit model selection is required.