定了，GPT-5.6 本周全量上线

On July 8, OpenAI announced that the GPT-5.6 family would go fully live on Thursday of that week (early Friday Beijing time). The family includes three models, named after celestial bodies: Sol (the Sun), Terra (Earth/Land), and Luna (the Moon). Sol is the flagship, aimed at complex reasoning, coding, scientific research, and cybersecurity; Terra is positioned as the daily workhorse with performance comparable to GPT-5.5 but at half the price; Luna focuses on speed and low cost, suitable for summarization, customer service, and batch content processing.

OpenAI CEO Sam Altman stated that Sol is significantly more efficient than previous versions on coding tasks, citing third-party benchmarks showing Sol outperforming Anthropic's Fable 5 and Opus 4.8. Notably, Anthropic's Fable 5 was originally scheduled for release on July 7 but announced a five-day delay to July 12 just hours before OpenAI's official announcement. This timing has been interpreted by outsiders as OpenAI deliberately capitalizing on a competitor's window of absence.

Meanwhile, OpenAI launched ChatGPT Work, an intelligent assistant for office scenarios that can draft documents, spreadsheets, and presentations. Microsoft also quickly announced that GPT-5.6 would become the 'preferred model' for Microsoft 365 Copilot, covering Word, Excel, PowerPoint, and Cowork. This move is seen as a response to rumors of cracks in the partnership — earlier reports suggested Microsoft was scaling up its own MAI models to reduce costs.

On the same day as the model release, OpenAI Applications CEO Fidji Simo announced she would step down from her full-time role for medical reasons and become a part-time advisor. Simo joined the OpenAI board in 2024 and took on the role in 2025, overseeing the company's commercial and product operations, with COO Brad Lightcap, CFO Sarah Friar, and former CPO Kevin Weil reporting to her. Her departure leaves a leadership vacuum, especially as OpenAI considers going public. Chief Revenue Officer Denise Dresser (former Slack CEO) is seen as a candidate for expanded authority. Altman publicly thanked Simo for her contributions.

That same week, OpenAI confirmed it is shutting down the ChatGPT browser Atlas, launched last October, integrating its intelligent browsing capabilities into a new Chrome extension and an upgraded ChatGPT desktop app. This strategic contraction comes amid heated competition in the AI browser space: Perplexity's Comet and The Browser Company's Dia are actively vying for market share. Previously, OpenAI had scaled back side projects like the video tool Sora due to internal pressure.

OpenAI faced a major legal setback in its copyright lawsuit. The New York Times and Daily News accused OpenAI of concealing information during the discovery phase: the plaintiffs claim OpenAI has already searched its training data and built a database containing tens of millions of de-identified ChatGPT conversations to assess potential infringement. The plaintiffs' lead attorney, Ian B. Crosby, said OpenAI's actions indicate it knows it is infringing and asked the court to sanction OpenAI and prohibit it from using what are considered unreliable chat log samples.

OpenAI spokesperson Drew Pusateri denied the allegations, calling the plaintiffs' request a violation of user privacy. The case had already drawn attention due to the Times' demand for the destruction of training data involving copyrighted material. The core of this battle: whether OpenAI has retained internal data sufficient to prove infringement within compliance boundaries, and whether the court will impose sanctions that limit OpenAI's use of evidence in future litigation. The outcome could have deep implications for the power balance between AI companies and content rights holders.

Behind the successive launches and delays of GPT-5.6 and Anthropic's Fable 5 lies a lack of a unified federal framework for evaluating frontier models. Mina Narayanan of Georgetown University and Databricks co-founder Andy Konwinski point out that the public and regulators lack transparent insight into how these models obtain clearance. This ambiguity fuels debate about the industry's influence over AI policy.

In just one week, OpenAI demonstrated its ability to juggle product, personnel, legal, and strategic challenges. If the tiered strategy of GPT-5.6 succeeds, it will strengthen OpenAI's coverage from high-end to mass markets, but internal leadership turbulence, legal risks, and regulatory uncertainty could continue to drag on its stable expansion. For most enterprise users, the real test is not how powerful Sol is, but whether Terra and Luna can be embedded into daily workflows at acceptable costs.