Google DeepMind confirmed on September 23 that Gemini 4 post-training is underway, with chief executive Koray Kavukcuoglu telling The Information’s AI Agenda Live Summit that the next flagship is being targeted for release “much earlier” than the end of 2026. Kavukcuoglu, who took over the lab on August 12, 2026, said early results justify shipping an aggressive frontier model rather than waiting for a polished one, a notable shift in tone for a research unit that historically has prioritized caution.
Pre-training on Gemini 4 started on July 21, 2026. The transition into Gemini 4 post-training inside roughly eight weeks is the fastest such handoff Google has executed, well ahead of the approximately 12 weeks recorded for the prior generation. The compressed schedule comes as Gemini 3.5 Pro has remained in partner testing since June 2026 after missing three separate launch deadlines, deepening pressure on Mountain View to ship a competitive model before the calendar flips.
Why Gemini 4 Post-Training Is Racing the Clock
The urgency behind Gemini 4 post-training is visible in the latest Intelligence Index v4.3.2 readings. Google’s current frontier model, Gemini 3.6 Flash, scores 34.0, while Anthropic’s Opus 5.5 leads at 57.6. OpenAI’s GPT-6 Astra sits at 52.7 and GPT-6 Sol at 47.5. A 24-point gap on a frontier benchmark is the difference between competing at the top of the market and competing one tier below it, and Kavukcuoglu described that gap as motivation rather than crisis during his September 23 appearance.
Kavukcuoglu said he is “a hundred percent confident” in the team and pointed to the “insanely strong” researchers still in place. The framing matters because internal stability at DeepMind has frayed. Noam Shazeer left for OpenAI in June 2026, AlphaFold laureate John Jumper joined Anthropic two days later, and Jeff Dean departed to co-found Discovery Loop with Oriol Vinyals, Quoc Le, and Sanjay Ghemawat. Researchers Jonas Adler and Alexander Pritzel also exited for Anthropic, leaving new-hire-to-departure ratios compressed from roughly 12 to 1 in Q2 2023 to about 2 to 1 in Q3 2026.
Gemini 4 Post-Training and the Pricing Reset
The Gemini 4 post-training push lands days after the seven-capability pricing structure collapsed. On September 12, Anthropic CEO Dario Amodei published “We Must Pace the Frontier,” urging frontier labs to slow capability releases so safety work could keep pace. Sam Altman signed on publicly, Elon Musk endorsed the direction, and departing DeepMind leader Demis Hassabis called the framing right while reserving judgment on details. Ten days later, Anthropic shipped Opus 5.5, and OpenAI followed 90 minutes later with GPT-6 Sol and Luna at permanent 50% price cuts. The coordinated pricing framework tied to Amodei’s call did not survive the week.
For Google, that breakdown is an opening. GPT-6 Luna now lists at $0.10 per million input tokens and $0.50 per million output, GPT-6 Sol at $2 and $10, Claude Opus 5.5 at $4 and $20, Claude Fable 5.1 at $10 and $50, and GPT-6 Astra at $10 and $50. Grok 4.7 arrived Monday at $2 and $6. Anthropic reports Opus 5.5 generates output 30 percent faster than Opus 5 with 40 percent lower workload cost, putting more pressure on Gemini 4 post-training to land a competitive intelligence score at a competitive price, not just one or the other.
What Gemini 4 Post-Training Says About Google’s Strategy
Kavukcuoglu used his first public appearance to recast DeepMind’s mission. “The conversation of [whether] we achieved AGI or not is not the right conversation,” he told The Information. “The conversation is more about are we able to build intelligent agents that we can trust?” The line marks a visible pivot from his predecessor’s farewell memo, in which Hassabis wrote that artificial general intelligence “is close at hand.” Reframing the lab around trusted agents rather than AGI arrival gives Google political cover to ship Gemini 4 on an accelerated schedule without triggering the safety arguments that followed the “Pace the Frontier” open letter signed by more than 1,200 frontier-lab employees in July.
According to The Guardian, one former Google executive put it bluntly: “The era of DeepMind as a standalone research actor is over.” Kavukcuoglu appears to be answering that critique by tying Gemini 4 post-training to a deliverable product rather than a research milestone. If Gemini 4 ships in November with credible capability numbers, the “Google has fallen behind” thesis loses its cleanest supporting data point and the three-way race between OpenAI, Anthropic, and DeepMind tightens before any of them finalize 2027 roadmaps.
The strategic implication running underneath all of this is that frontier pricing will now reset roughly every eight weeks rather than every six months, and any lab still operating on a research-first calendar will be quoting yesterday’s benchmarks. Gemini 4 post-training is Google’s formal entry into that cadence, and the next test will be whether the Intelligence Index v4.4 readings in late October show a model inside striking distance of Opus 5.5 or another entry still hovering near the 34-point line.
Source: the-decoder.com