Kavukcuoglu: Gemini 4 hits post-training, early ship eyed

Google's disclosure arrived not through a scheduled product announcement but through an interview: Koray Kavukcuoglu discussed Gemini 4 with The Information at its AI Agenda Live Summit on September 23. The article notes that Google's communications team has stayed quiet since, and that no parameter count, training cost, pricing, or benchmark scores have been published.

That reticence stands against two rival plays described in the body. OpenAI disclosed a 100,000-GPU training run for GPT-6 Astra and president Greg Brockman's internal 'AGI era' framing. Anthropic cut Claude Opus 5.5 pricing 20% and cache costs 60% at launch. Google's own history is also relevant: after ChatGPT's late-2022 debut it issued an internal 'code red' and pushed out Bard within months, a launch widely criticized for factual errors in its own demo. Gemini 3 already competes credibly with GPT-6 and Claude Opus 5.5 on most public benchmarks, so the article reads an early Gemini 4 as an attempt to extend existing strengths in multimodal reasoning and distribution rather than to close a gap from zero.

What the outcome hinges on, per the body, is whether Gemini 4 clears a bar on tasks enterprises actually run in production, and whether its price matches rival moves that have already compressed margins. The article also flags that Anthropic sells through Google Cloud, so Google hosts a rival's flagship model on its own infrastructure while trying to beat it. Any Gemini 4-driven shift in cloud share is expected to lag the launch by a quarter or more, given enterprise procurement cycles, meaning the more telling signal may be whether Google leads with specs and pricing at launch rather than a timeline.