Koray Kavukcuoglu, elevated around 12 August 2026 as Google DeepMind’s top AI executive after Demis Hassabis stepped back from day-to-day AI leadership, used The Information AI Agenda Live summit (23–24 September) to update Gemini 4’s status: the model is already in early post-training.
Kavukcuoglu said Google intends to release «much earlier» than waiting for the end of 2026, prioritizing a useful early post-training output as soon as practical. The message corrects expectations of a drawn-out cycle and puts Google in catch-up mode against rivals advertising next-generation milestones.
Gemini 4’s pre-training run was announced around 21 July 2026. Reaching post-training within roughly two months suggests the base run cleared internal gates for alignment, distillation, and preference tuning—without disclosing public quality metrics.
Google’s last flagship family, Gemini 3, debuted in November 2025. Since then, Gemini 3.5 Pro has been described as stuck in partner testing while the company pushed Flash variants for latency and cost. That mix—slow flagship, aggressive Flash—left a perceptual gap in the leadership narrative.
September 2026’s competitive board includes GPT-6 Sol/Luna lines, Claude Opus 5.5, and Grok 4.7, each compressing media attention. Google needs Gemini 4 not to arrive as a year-end footnote.
Some recent coverage placed Google «about 40% behind» leaders on an Intelligence Index. Kavukcuoglu did not treat that figure as dogma; Neural Pulse cites it only as third-party reporting, not as a measurement verified in this piece.
The leadership shift matters. Hassabis retains scientific and AGI-vision weight, but Kavukcuoglu’s elevation concentrates product execution and release calendars. Investors read the change as a bet on shipping speed without abandoning DeepMind’s research brand.
Early post-training is not general availability. Red-teaming rounds, safety evaluations, and capacity talks with Cloud and Android still await. Even so, the public signal reduces uncertainty over whether Gemini 4 was merely a roadmap rumor.
For enterprise developers, the practical questions are API compatibility and pricing. If Google accelerates the flagship, it must clarify 3.5 Pro’s partner path and whether Flash remains the default on-ramp or cedes the spotlight.
Hardware analysts noted that post-training pace also depends on HBM clusters and internal inference queues. Speeding the calendar means prioritizing GPUs over other Cloud products—an internal political decision as much as a technical one.
Rivals will read any further slip as confirmation of the lag narrative; any early demo as proof that the coverage’s «40%» was a snapshot, not a destiny. Kavukcuoglu’s communication aims to move that perception before a version number hits the public site.
Neural Pulse keeps the dated facts: Kavukcuoglu’s August elevation, July pre-training announcement, late-September early post-training confirmation, and ambition for a release ahead of a year-end horizon.
What is not on the public record—evaluation thresholds, an exact GA date, or an internal ranking versus GPT-6 or Opus 5.5—must await product notes or papers. Until then, this is a calendar-and-leadership story, not a closed benchmark.
In a market where each quarter redefines the perceived state of the art, Google is betting Gemini 4 arrives soon enough to compete in the same conversation, not the next one.
