OpenAI might have inadvertently saved Google's life. That's the provocative take from Yao Shunyu, a former Anthropic research scientist now at Google DeepMind, who sat down for a podcast called "Language is the World" to dissect how Gemini finally turned the tables on ChatGPT.
Yao argues that Gemini's market resurgence came down to two things hitting in sequence: first, a viral image-generation product called Nano Banana that drove a flood of app downloads, followed by the launch of Gemini 3, which turned those downloads into loyal users. Third-party data backs him up — by early 2026, Gemini's app market share had climbed to around 25%, while ChatGPT's had slipped from 69% to 45%.
But the deeper story is about timing. The industry once feared that chatbots would eat search alive, Yao explains. OpenAI got there first, but didn't finish the job — it didn't swallow search whole. That gave Google the breathing room it needed to catch up. And once Google's own chatbot reached parity, the pressure shifted back onto OpenAI.
Inside Google, the development playbook has changed too. Pre-training, Yao says, has become a comfortable, top-down engineering project — Google's operational prowess is a perfect fit. Recent updates to the Gemini API, including multimodal file search and page-level citations, have further extended Gemini's reach. Meanwhile, post-training, with its higher uncertainty, remains a bottom-up research experiment.
So what makes Gemini 3's long-context performance so strong? Yao mentioned a trick from the pre-training phase that surprised him — then smiled and declined to share details.
Google is also pushing rapid developer updates — a breaking change to the Gemini API is coming, giving developers until June 8 to update their code.