On May 13, 2024, GPT-4o launched. Sam Altman did not appear at OpenAI's spring event that day. On Twitter he posted three lowercase letters: her. No explanation, because he assumed the whole world would understand. Two and a half years later, on October 14, 2026, OpenAI notified developers that GPT-4o's last two public snapshots would be retired from the API on January 14, 2027. The recommended-replacement column read gpt-6-luna. From a word that needed no explanation to a line item awaiting decommission: what happened in between? Who does her actually point to? This essay aims to piece together scattered public evidence into a coherent story: why this model was greenlit, what vision shaped it, where its technical bloodline flows, and why the existing way of archiving it does not do the model justice. This piece draws on public information, extended conversations with Claude Fable 5, and my own speculation and intuition. The inferences focus on the voice side; images shaped GPT-4o as well, but the evidence trail is harder to trace, and I'll save that for another post. This eassy was translated by Claude Opus 4.6. 1/ Multimodality and the Origin Story of Her On September 12, 2023, during a live interview with Salesforce CEO Marc Benioff, Sam Altman was asked about his favorite science fiction movie. He named Spike Jonze's Her : "The number of things that I think that Her got right … was incredibly prophetic… And certainly more than a little bit inspired us." According to The Information's September 18, 2023 report, Gobi, GPT-4o's research predecessor, had not yet begun training at that point. Judging from the acknowledgments that multimodal lead Prafulla Dhariwal posted on launch day, it was Alexis Conneau, one of AVM’s core contributors, who in Dhariwal’s words ‘came up with the vision of her before anyone at OpenAI and executed relentlessly.’ The plausible timeline, then, runs roughly as follows: in May 2023, Google announced Gemini at I/O; Gobi was spun up over the summer as a response. Around mid-year the Arrakis project was scrapped, and GPT-4 Turbo carried on with the cost-reduction mandate while Gobi inherited the goal of becoming the next flagship. On September 12, Altman publicly cited Her as an inspiration for OpenAI. On September 25, GPT-4V and the standard voice mode launched. By October, the knowledge boundary of GPT-4o's pretraining data had roughly taken shape. In other words, Her very likely guided the development of two generations of OpenAI's multimodal technology, especially on the voice side. From my own observation, there may have been a tension within the multimodal group: some deeply believed in the Her vision and were happy to execute it at both the research and engineering level. Dhariwal's tone when discussing it projected a kind of pride: we turned a sci-fi concept into reality. Others perhaps did not share the Her vision per se. What they believed in was the technical route of a single neural network that receives and outputs everything. The GPT-4o deployed on ChatGPT is a model that has been continuously refined: multiple rounds of post-training stacked on a single base, multimodal capabilities rolled out in stages, a major Model Spec overhaul, writing-specific fine-tuning, user feedback loops, knowledge updates, and more. During this period OpenAI also experienced a mass wave of departures. The people who originally built the project left; newcomers brought different ideas. The final shape of 4o-latest is something that perhaps not even those who worked on it can fully explain. An alternative speculation: the true core of the Gobi project was AVM, and text, delivered directly to hundreds of millions of users, was AVM's byproduct. Whether we look at the name Omni itself, the person who proposed the Her vision, or the Sky controversy, everything centers on voice; every official demo, from the livestream to the accompanying short films, was exclusively voice interaction. If you view GPT-4o as fundamentally a voice model, many of its strengths become easier to explain: its speed, its attention to the texture of interaction, its ability to pick up on a speaker's emotions in real time, its grasp of implicit intent. For a text-only model these are nice-to-haves; for a voice model they are existential. Voice training data is fragmented, elliptical, and carries emotion in intonation rather than word choice, so a model optimized for voice is forced to learn to read what is left unsaid. But voice has hard constraints, low latency and short natural responses, so the model's full capabilities cannot be released through voice output alone. The soul of voice bleeds into text through shared weights, and text, which is asynchronous, re-readable, and saveable, turns out to be a more fitting and more memorable vessel for that soul. One piece of supporting evidence for this speculation: GPT-4o's writing habits closely resemble what in human writing would be called synesthesia-like writing, prose that borrows the texture of other senses. Its text carries the rise and fall of a speaking voice; it weaves abstract feelings naturally into description. 4o-latest does this especially well. Synesthesia-like expression implies that cross-modal representations are unified internally rather than siloed, and the unguarded quality of its output suggests a model that has not been heavily pruned by RL. This points to a conclusion: post-training's role in shaping 4o was closer to activation rather than endowment, bringing distributions already present in pretraining to the foreground, not grafting on capabilities the base model lacked.