OpenAI announced on 10 August 2021 that Codex, a descendant of GPT-3 fine-tuned on publicly available code, would power GitHub Copilot and be made available via API in private beta [source]. The model was trained on tens of millions of public repositories and demonstrated proficiency in Python, with additional capability in over a dozen languages including JavaScript, Go, Perl, PHP, Ruby, Swift, and TypeScript.

According to the announcement, Codex could interpret simple commands in natural language and execute them on behalf of the user, enabling tasks such as transpilation, code explanation, and refactoring. OpenAI reported that Codex solved 37% of problems correctly on the first attempt in an internal evaluation, and that allowing 100 submissions per problem raised the success rate to 77%.

The company acknowledged safety considerations, noting that Codex could generate code with subtle bugs or exploitable vulnerabilities. OpenAI stated it had implemented safety measures including output filtering and monitoring for misuse, and that the model's suggestions should be treated as any code from an external source—carefully reviewed and tested before use.

The announcement marked a shift in OpenAI's product strategy toward domain-specific applications of large language models. Codex represented one of the first commercially deployed code-generation systems built on transformer architectures at scale.

Developers were invited to join a waitlist for API access. OpenAI indicated that feedback from the private beta would inform future development and safety mitigations. The model's release followed earlier demonstrations of GPT-3's limited code-writing ability and preceded broader industry adoption of AI-assisted programming tools.