Codex runs jobs unattended; Altman lokes ‘Goblin’ model

OpenAI CEO Sam Altman wrote on X that he started a batch of Codex coding jobs, spent time with his child, and returned at naptime to find the tasks completed. He added the result made him “very optimistic for the future.”

Codex converts natural-language prompts into working code and can be configured to hold a list of tasks, sequence steps and return finished output without a developer reviewing each intermediate action. OpenAI describes the tool as an autonomous agent rather than a simple autocomplete feature.

The company has highlighted these hands-off capabilities in pitches to enterprise customers and expanded its cloud offerings in partnership with Microsoft. Competing firms, including Anthropic and Google, are developing unattended coding assistants with similar end-to-end automation goals.

Altman ran a poll on X asking users what they wanted in the next model. Many responses requested “more goblins,” and he later wrote that naming the next release “Goblin” was “almost worth it to make you all happy.”

On April 29, OpenAI published a report titled “Where the Goblins Came From.” The report examined why several recent model versions began using metaphors involving goblins, gremlins and similar creatures. It attributed the pattern to multiple small incentives during training, including work on personality customization that rewarded a “nerdy” conversational style. The company traced the shift in metaphor use to model versions beginning around GPT-5.1.

Earlier comments from Altman described the model as an “autistic genius,” a phrasing he used on X to characterize a system with strong technical abilities and variable performance on tone and wording.

OpenAI has said it will release further models and updates. The company’s public posts and technical report provide the timeline and rationale for the changes to model behavior and product positioning.