OpenAI published a blog post on 5 January 2025 arguing against open-weight AI models, stating that "open weights" do not guarantee safety or accessibility and may concentrate power among well-resourced actors [source].

The post claims that open-weight models require significant compute infrastructure to run effectively, limiting their practical availability to organisations with substantial resources. OpenAI asserts that its API-based approach provides broader access than downloadable model weights, citing lower barriers to entry for developers without hardware investments.

The company positions this argument within ongoing policy debates about AI model distribution. OpenAI contends that releasing model weights creates security risks, including potential misuse for generating harmful content or enabling malicious applications, without delivering the democratisation benefits proponents claim.

The post does not announce changes to OpenAI's existing model release practices. GPT-4 and subsequent flagship models remain available exclusively through API access, with no model weights released publicly. OpenAI has not released open-weight versions of any production models since GPT-2 in 2019.

The timing coincides with regulatory discussions in multiple jurisdictions about mandatory open-weight requirements for certain AI systems. OpenAI's position directly opposes proposals from open-source advocates and some policymakers who argue that publicly available weights enable independent safety research and reduce vendor lock-in.

The post frames OpenAI's closed-weight approach as aligned with safety and accessibility goals, stating that API distribution allows the company to implement usage policies and monitor for abuse. No technical evidence or comparative data on access patterns between open-weight and API-only models is provided in the blog post.