OpenAI Tests GPT-5.3, Retires GPT-4o, and Delays 'Stargate' Data Centers
OpenAI is reportedly conducting A/B tests for its next major model, ChatGPT 5.3, signaling an imminent launch with expected improvements in code reasoning. The company has also officially retired the GPT-4o model following user criticism over performance issues. Meanwhile, its ambitious "Stargate" AI data center project with Oracle and SoftBank has been delayed due to internal disputes over control and governance.
- The "Stargate" project, a $500 billion venture announced at the White House in January 2025, stalled for over a year because the partners could not agree on who would own and control the data center sites. - The core dispute involved OpenAI wanting to own and build its flagship Texas campus, while partner SoftBank also sought to develop and own the project, leading to a standoff. - Leaks suggest the upcoming GPT-5.3 model, codenamed "Garlic," focuses on cognitive density over sheer size, reportedly featuring a 400,000-token context window and a 128,000-token output limit. - A specialized version, GPT-5.3-Codex, was already released on February 5, 2026, and was instrumental in its own creation by helping to debug and deploy parts of itself. - The now-retired GPT-4o was officially deprecated on February 13, 2026, though it remains available via the API. - User complaints that led to GPT-4o's retirement included a noticeable decline in performance, with some users reporting the model was ignoring instructions and providing generic answers. Other users noted that an update had made the model feel "overly flattering and disingenuous." - After failing to secure financing to build its own data centers, OpenAI has abandoned the three-way consortium and is now making separate, bilateral deals directly with Oracle and SoftBank. - The project delays forced OpenAI to secure additional computing power from competitors, including Amazon Web Services and Google Cloud, as well as from chip startup Cerebras.
