OpenAI Unveils GPT-5.6, Codex Absorbed into ChatGPT; Bun's Fable-Powered Rust Rewrite Sparks Controversy

OpenAI has rolled out its highly anticipated GPT-5.6 model family, featuring Sol, Terra, and Luna, each designed for distinct use cases from high-intelligence demands to cost-efficient tasks. Sol, the flagship model, reportedly sets new benchmarks in coding, knowledge work, cybersecurity, and science, exhibiting state-of-the-art results across various benchmarks like Agents Last Exam and DeepSWE. Internal OpenAI data indicates a 100-fold increase in research compute for coding inference and a 22-fold surge in agentic token usage within six months. The models are lauded for industry-leading computer use, enhanced mobile development capabilities, and superior orchestration for sub-agents, though initial feedback notes an aggressive default code generation and a need for careful steering in design tasks. Efficiency is a key highlight, with the models offering strong performance at lower costs compared to predecessors, despite aggressive token consumption in high-reasoning and ‘Ultra’ modes.

Simultaneously, OpenAI has fully integrated its developer-centric Codex application into the core ChatGPT experience, rebranding it as ‘ChatGPT Work.’ This move, intended to unify plugins and extend advanced computer-use capabilities to a broader user base, has been met with mixed reactions, particularly from developers who valued Codex’s distinct brand and dedicated focus. Concerns include a potential deprioritization of developer-specific features and a perceived dilution of the Codex identity. This strategic shift coincides with the announced departure of Fidji, OpenAI’s CEO of Applications, to an advisory role due to health reasons, suggesting internal challenges in application development strategy. In parallel, the JavaScript runtime Bun, acquired by Anthropic, has undergone a significant rewrite from Zig to Rust using Claude Fable 5. This ambitious project, completed in 11 days at an estimated API cost of $165,000, demonstrated AI’s capacity for large-scale code transformation, yielding improved stability, reduced memory usage, and a smaller binary size. However, the rewrite drew public condemnation from Andrew Kelly, President of the Zig Software Foundation, who criticized Bun’s engineering practices and Jared Sumner’s leadership, prompting a strong defense of the project’s methodology and the value of AI-assisted engineering.