Google DeepMind says Gemini 4 has entered post-training, launch 'much earlier' than year-end

In his first public media appearance since being elevated to lead Google DeepMind, Koray Kavukcuoglu said Gemini 4 is now in early post-training — the phase where a base model is refined for behavior and run through safety tests — and that Google intends to release an early version 'as soon as possible' because internal results are promising. The subtext is competitive urgency: Google has not shipped a flagship since Gemini 3 in November 2025, an unusually long gap in a year where rivals have iterated rapidly.

On the numbers cited, Gemini 4 reportedly trails Anthropic's Opus 5.5 (57.6 on the referenced benchmark) but leads OpenAI's GPT-6 Astra (52.7) and the new Sol (47.5) — a positioning that would put Google back in serious contention if it ships soon. Google is already dogfooding Gemini 4 internally to power its Antigravity coding tool.

Separately Google is not standing still on the current line: it debuted Gemini 3.8 text-to-speech with voice replication, underscoring a steady cadence of incremental releases while the flagship finishes.

The skeptical read, voiced by developers in the sources, is timeline fatigue: 'much earlier than year-end' is a soft commitment, and Google has slipped Gemini timelines before. With OpenAI and Anthropic having just cut prices aggressively, Gemini 4 will need both competitive benchmarks and competitive economics to matter. Watch for a concrete ship date and whether the 'early version' is a limited preview or a full API launch.